Robust Output Feedback Stabilization,
Compressors Surge and Stall Example
by
Ali Rajaeesani
A thesis submitted in conformity with the requirementsfor the degree of Masters of Applied Science
Graduate Department of Electrical and Computer Engineering
University of Toronto
Copyright c© 2003 by Ali Rajaeesani
For My Father
Abstract
Robust Output Feedback Stabilization,
Compressors Surge and Stall Example
Ali Rajaeesani
Masters of Applied Science
Graduate Department of Electrical and Computer Engineering
University of Toronto
2003
This thesis introduces an approach to output feedback stabilization of SISO nonlinear
systems which are not uniformly completely observable and are affected by disturbances.
This approach is applied to the output feedback control of axial flow compressors in
jet engines represented by a three state model developed by Moore and Greitzer. The
extension of this approach to the case when constant unknown parameters affect the
plant is addressed and its limitations are discussed.
ii
Acknowledgements
I would like to thank Manfredi Maggiore, my supervisor, for his guidance and many
suggestions, which were crucial to the successful completion of this project.
I am grateful to Mohammad and Majid, my uncles, and Ali Nezeran for their constant
support during my studies in Canada. Without them this work would never have come
into existence (literally).
I wish to thank my family for their patience and love, especially my brother for his
constant support and encouragement.
Finally, I would like to thank the following: Samoodis (for their support through the
early years of chaos and confusion); Payam, Amy, Mehran and Amir (for their friendship);
Arash, Ida, Ashkan, Negar, Farhad, Mohammadreza, babak ... (for all the good times
we had together); My wonderful friends in Systems Control Group, especially Maziar,
Rashid, Peyman, Amir and Omid (for sharing their knowledge and jokes); and my friends
in IAUT (they know why).
Toronto, July 2003 Ali Rajaee-Sani
iii
If the facts don’t fit the theory, change the facts.
-Albert Einstein
Contents
Contents vi
1 Introduction 1
1.1 Literature Review . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
1.1.1 Output Feedback Control . . . . . . . . . . . . . . . . . . . . . . 3
1.1.2 Surge and Stall . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
1.2 Organization of the Thesis . . . . . . . . . . . . . . . . . . . . . . . . . . 5
2 Disturbed Non-UCO Systems, the Theory 6
2.1 Problem Formulation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
2.2 Nonlinear Observer . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
2.3 Projection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20
2.3.1 Projection Sets . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
2.3.2 Observer Estimates Projection . . . . . . . . . . . . . . . . . . . . 26
2.4 Boundedness and Closed Loop Stability . . . . . . . . . . . . . . . . . . . 32
2.5 Concluding Remarks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35
3 Application to Axial Flow Compressors 37
3.1 Compressors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38
3.1.1 Types of Compressors . . . . . . . . . . . . . . . . . . . . . . . . 38
3.1.2 The Axial Compressor, Principles of Operation . . . . . . . . . . 38
v
3.1.3 Surge . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39
3.1.4 Rotating Stall . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41
3.1.5 Modelling of Axial Compression Systems . . . . . . . . . . . . . . 43
3.1.6 Disturbances . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 46
3.1.7 Change of Variables . . . . . . . . . . . . . . . . . . . . . . . . . . 47
3.2 Output Feedback Control for MG3 . . . . . . . . . . . . . . . . . . . . . 48
3.2.1 Observability Mappings . . . . . . . . . . . . . . . . . . . . . . . 48
3.2.2 State Feedback Control . . . . . . . . . . . . . . . . . . . . . . . . 50
3.2.3 Input-to-State Stability . . . . . . . . . . . . . . . . . . . . . . . . 53
3.2.4 Stabilizing Control Law for Augmented System . . . . . . . . . . 57
3.2.5 Observability Set . . . . . . . . . . . . . . . . . . . . . . . . . . . 60
3.2.6 Observer Design . . . . . . . . . . . . . . . . . . . . . . . . . . . . 62
3.2.7 Simulation Results . . . . . . . . . . . . . . . . . . . . . . . . . . 62
3.3 Concluding Remarks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63
4 Observability of Systems with Uncertainty 68
4.1 MG3 with Uncertainty . . . . . . . . . . . . . . . . . . . . . . . . . . . . 68
4.2 Controller Design . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 69
4.3 Observability Map . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 74
4.4 Uncertainty and Observability Map . . . . . . . . . . . . . . . . . . . . . 75
5 Conclusion 87
Bibliography 89
vi
List of Figures
2.1 Block diagram. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
2.2 Projection mechanism. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
2.3 Properties of C. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
2.4 Properties of C,cont. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
2.5 Geometric Interpretation of the projection condition. . . . . . . . . . . . 27
2.6 G Transformation. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
3.1 Blade rows. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39
3.2 Compressor characteristic. . . . . . . . . . . . . . . . . . . . . . . . . . . 40
3.3 Compressor characteristic with surge and surge avoidance lines. . . . . . 41
3.4 Shematic drawing of hysteresis caused by rotating stall. . . . . . . . . . . 42
3.5 Physical mechanism for inception of rotating stall. . . . . . . . . . . . . . 43
3.6 Compressor system. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 44
3.7 Cubic compressor characteristic of Moore and Greitzer. . . . . . . . . . . 46
3.8 Sinusoidal disturbances. . . . . . . . . . . . . . . . . . . . . . . . . . . . 64
3.9 Constant disturbances with different values of ρ. . . . . . . . . . . . . . . 65
3.10 Decaying exponential disturbances . . . . . . . . . . . . . . . . . . . . . 66
3.11 Peaking phenomenon. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 67
4.1 MG3 with adaptive controller. . . . . . . . . . . . . . . . . . . . . . . . . 73
4.2 x and its estimate, Example 4.4.1. . . . . . . . . . . . . . . . . . . . . . . 77
vii
4.3 Projection effect, Example 4.4.1. . . . . . . . . . . . . . . . . . . . . . . . 78
4.4 x and its estimate, Example 4.4.2. . . . . . . . . . . . . . . . . . . . . . . 82
4.5 Projection effect in x coordinates, Example 4.4.2. . . . . . . . . . . . . . 83
4.6 Projection effect in ye coordinates, Example 4.4.2. . . . . . . . . . . . . . 84
4.7 Disturbance and uncertainty, Example 4.4.2. . . . . . . . . . . . . . . . . 85
4.8 Disturbance and uncertainty, projection effect, Example 4.4.2. . . . . . . 86
viii
Chapter 1
Introduction
Two of the fundamental instability phenomena in axial flow compressors are surge and
rotating stall. Rotating stall develops when there is a region of stagnant flow rotating
around the circumference of the compressor which causes undesired vibrations in the
blades and reduces pressure rise of the compressor and surge is an axisymmetric oscillation
of the flow through the compressor that can cause damage and undesired vibrations in
other components of the system [23]. Moore and Greitzer in [26] developed a model for
axial flow compressors which is frequently used in the control literature.
The Moore-Greitzer model has three states which are rotating stall, mass flow and pres-
sure rise (see Section 3.1.5). One of the applications of axial air flow compressors is in jet
engines. Commercial jet engines include temperature, pressure and rotor speed sensors,
but most often do not include sensors for mass flow, hence making it desirable to esti-
mate. This, together with the fact that sensors for stall are not commercially available,
makes the problem of controlling stall and surge inherently an output feedback control
problem, whereby the only measurable variable is the pressure rise. One of the features
of Moore-Greitzer three state model (MG3) is that when pressure rise is the only mea-
surable state, the model becomes unobservable whenever there is no flow through the
compressor; in other words it is observable on an open region of the state space, rather
1
Chapter 1. Introduction 2
than everywhere, or it is not uniformly completely observable (Non-UCO). This is one of
the reasons that most of the researchers have only focused on the development of state
feedback controllers for this model which may not be implementable since they assume
that all of the states of this model are measurable. To the best of our knowledge, output
feedback control solutions using only pressure rise as the measurable state in MG3 do not
rely on the estimation of the entire state of the system, and, except in [23], no attempt
has been made to solve this problem. In [23], the authors apply the observer that they
have developed in [16] to estimate the entire state based on the pressure rise as the only
measurable state.
A missing feature of the work in [23] is the absence of any external disturbances or
uncertainties in the MG3 model they use, which is the motivation of this thesis. We first
tried to apply the observer developed in [16] to an MG3 model subject to some general
external disturbances. In our path we realized that the technique we were developing for
this particular MG3 problem could be applied to a class of non-UCO nonlinear systems
whose observability maps have a particular specification (see Assumption A1) , and are
subject to some slow varying or constant external disturbances. This observation led
us to develop a more general theory for this class of systems which also includes the
MG3 model. This theory is illustrated in Chapter 2. As the next natural extension, we
considered the case when the MG3 model is affected by constant unknown disturbances.
Our attempts in this direction were not successful and we realized that the observability
mapping in MG3 and systems with uncertainties similar to the MG3 model has a feature
(see Section 4.4) that prevents us from using the observer in [16] with its current format.
This line of work needs further investigation which is beyond the scope of this thesis. As
a result we have not developed a general theory for non-UCO systems with uncertainties,
but again we show that the observer in [16] can be applied to a class of nonlinear non-UCO
systems whose observability mappings have a particular feature (see Section 4.4).
Chapter 1. Introduction 3
1.1 Literature Review
In what follows we present a literature review both on nonlinear output feedback sta-
bilization and the surge and rotating stall control problem and later we outline the
organization of this thesis. We emphasize that the MG3 application part of this thesis is
purely theoretical and is therefore to be seen as an application of the theory developed
in Chapter 2.
1.1.1 Output Feedback Control
In 1992 H. Khalil and F. Esfandiari published their work [1] on nonlinear output feed-
back area where they introduced a new technique for designing robust output feedback
controllers for input-output linearizable systems. Their technique has two features. First
a high-gain observer that estimates the derivatives of the output and second, a bounded
state feedback control, obtained by saturating a continuous state feedback function out-
side of a compact set, and hence protecting the state of the plant from peaking when
the high gain observer estimates are used. This technique has been adopted by several
researchers in their work (see e.g. [2, 3, 4, 5, 6, 7]) to solve various problems in output
feedback control. In most of these papers the authors consider input-output feedback
linearizable systems. One of the remarkable works in this line is [9], which proves a
separation principle for a rather general class of nonlinear systems.
In most of the existing and mentioned works in the literature on output feedback control,
the controller is designed in two steps. First, a globally bounded state feedback control
is designed. Second, a high-gain, fast observer, recovers the performance achieved by the
state feedback. Most of these works have two common features which are, the assumption
of uniformly complete observability, and using the vector [y, y, . . . , y(ny), u, u, . . . , u(nu)]>
as feedback, for some integers ny, nu, where y and u denote the system output and
input, respectively. The latter feature implies that, for systems which are not input-
Chapter 1. Introduction 4
output feedback linearizable, the explicit inverse of the observability mapping has to be
known. Both of the mentioned features have the disadvantage that in some situations
systems may not possess either or both of these two properties.
In [16], which this thesis is based on, the authors achieve a separation principle for systems
that are not observable on some regions of the state space and the input space, or in
short, systems that are not uniformly completely observable (non-UCO). The interesting
feature of their work is that for implementation of the controller, they do not need the
knowledge of the explicit inverse of the observability mapping. As we will show, these
two properties are essential for our work on the MG3 model.
1.1.2 Surge and Stall
In 1986 Moore and Greitzer developed a three-state model (MG3) for axial flow com-
pressors which became a benchmark in the control literature for several researchers (e.g.,
[24, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37]) to design stabilizing controllers for stall
and surge. As we mentioned earlier, most of the existing results, in some cases robust
ones (e.g. see [24]), assume that all the states are measurable and the focus is on the
development of state feedback controllers, which may not be implementable, and in the
area of output feedback control, where compressor pressure rise is the only output, not
all the system states are estimated (see e.g. Sections 12.6 and 12.7 in [40]). For a rather
complete discussion on the literature and existing models of axial compressors and meth-
ods for controlling surge and stall phenomena, one can refer to [39]. The work in [23] is
the only one which addresses the output feedback control problem for MG3 by estimating
the entire state of the system. However, a limitation of [23] is that no disturbances affect
the model.
As in all physical systems, disturbances occur in compression systems. Authors in [17]
and [18] study the effect of circumferential inlet flow distortion on the stability properties.
Chapter 1. Introduction 5
In [19] and [20] mass flow and pressure disturbances are studied. Authors in [21] study the
effects of inlet temperature distortion and find that these effects can be as significant as
those of inlet pressure distortion. As in [19], the effect of pressure and flow disturbances
are considered in this thesis (see Section 3.1.6).
1.2 Organization of the Thesis
In Chapter 2 we introduce the class of SISO systems affected by a particular kind of
disturbance which the technique in [16] can be applied to, and we develop the theory
which underlies our technique. In Chapter 3, after a brief discussion on the Moore-
Greitzer model for surge and stall in axial flow compressors, we apply our technique to
stabilize the MG3 model in the presence of disturbances. In Chapter 4 we discuss our
attempts to apply the technique in [16] to the MG3 model when there is uncertainty in
the compressor model, and after that we discuss the class of systems with uncertainty
that this output feedback stabilization technique can be applied to.
Chapter 2
Disturbed Non-UCO Systems, the
Theory
In this chapter we develop an output feedback control theory for a class of SISO, non-
UCO, nonlinear systems that are affected by some disturbances. First we describe the
problem formulation and our assumptions. Later we describe our nonlinear observer
structure and prove its properties. Finally we prove the closed loop stability when using
the observer estimates at the control input. Parts of this chapter rely on the work in [16].
2.1 Problem Formulation
Consider the following system,
x = f(x, u) + δ(t)
y = h(x, u)
(2.1)
where x(·), δ(·) ∈ Rn, δ(t) is a bounded disturbance with bounded time derivatives and
supt ‖δ(t)‖ = d, u, y ∈ R, f and h are known smooth functions and f(0, 0) = 0. In this
chapter we show that the closed-loop output feedback system depicted in the Figure 2.1
is uniformly ultimately bounded. Υ(xP , z, y) represents the observer (2.49), 0 ≤ nu ≤ n
6
Chapter 2. Disturbed Non-UCO Systems, the Theory 7
δ(t)
x = f(x, u)
y = h(x, u)
u
z
y
v = φ(xP , z)
Υ(xP , z, y)xP
nu
Figure 2.1: Block diagram.
is an integer and v = φ(xP , z) is the control law which is the input to a chain of nu
integrators. We also find an estimation of the domain of attraction of closed-loop system
and a bound on the disturbance δ(t).
We start by defining the following observability mapping H ,
ye4=
y
...
y(n−1)
=
h(x, u)
φ1(x, z)
...
φn−1(x, z)
+
0
θ1(δ)
...
θn−1(δ)
= H1(x, z) +H2(δ) = H(x, z, δ) (2.2)
(with a slight abuse of the notation, notice that ye is a vector in Rn and not the e − th
element of y) where z = (u, u, u, . . . , u(nu−1)), δ = (δ, δ, . . . , δ(nδ)), nu and nδ are integers
(0 ≤ nu ≤ n, 0 ≤ nδ ≤ n− 1), and
Chapter 2. Disturbed Non-UCO Systems, the Theory 8
y = ye,2 =∂h
∂xf(x, u) +
∂h
∂uu+
∂h
∂xδ(t) (2.3)
= φ1(x, z) + θ1(x, δ)
y = ye,3 =∂φ1(x, z)
∂xf(x, u) +
∂φ1(x, z)
∂zz +
∂φ1(x, z)
∂xδ(t) +
∂θ1(δ)
∂δ˙δ(t) (2.4)
= φ2(x, z) + θ2(x, z, δ)
...
y(i−1) = ye,i =∂φi−2(x, z)
∂xf(x, u) +
∂φi−2(x, z)
∂zz +
∂φi−2(x, z)
∂xδ(t) +
∂θi−2(δ)
∂δ˙δ(t) (2.5)
= φi−1(x, z) + θi−1(x, z, δ)
...
where
∂φi(x, z)
∂zz
4=
ni∑
k=1
∂φi∂zk
zk+1 (2.6)
Remark 1: In general θi’s are functions of x, z and δ, but in our case we assume that
θi’s are functions of only δ to get the form of mapping (2.2); see assumption A1.
The map (2.2) is parameterized by a vector z containing the derivatives of the control
input u, which for our purpose should be available for feedback. To this end, we augment
the system dynamics with nu integrators at the input side, which corresponds to using a
compensator of order nu, so that the state of the compensator gives precisely the desired
vector z. System (2.1) can be rewritten as follows,
x = f(x, z1) + δ(t), z1 = z2, . . . , znu= v. (2.7)
Define the extended state variable X = [x>, z>]> ∈ Rn+nu, and the associated extended
Chapter 2. Disturbed Non-UCO Systems, the Theory 9
system
X = fe(X) + gev + δe(t)
y = he(X)
(2.8)
where fe(X) = [f>(x, z1), z2 , . . . , znu, 0]>, ge = [0, . . . , 1]>, δe(t) = [δ(t), 0, · · · , 0]>,
and he(X) = h(x, z1). Now, we are ready to state our assumptions.
Assumption A1(Observability): Assume that ∀t ≥ 0, δ(t) ∈ ∆, where ∆ is a
compact set. Suppose O ⊂ Rn × R
nu which is an open set contains the origin, and
F1 : O → Rn × R
nu and F2 : O → Rn × R
nu are two smooth functions such that the
mapping (x, z) 7→ (ye, z) has the following structure,
Y =
ye
z
= F (x, z, δ)
4= F1(x, z) + F2(δ) =
H1(x, z)
z
+
H2(δ)
0
. (2.9)
Assume further that H1 is a diffeomorphism with respect to its first argument over O so
that, ∀Y ∈ Y 4= Y ∈ R
n+nu| Y = F1(x, z) + F2(δ), (x, z) ∈ O, δ ∈ ∆,
x
z
= F−1(Y, δ)
4= F−1
1 (Y − F2(δ)) =
H−11 (ye −H2(δ), z)
z
. (2.10)
In other words, we assume that as long as (x, z) ∈ O, having the output derivatives ye
and knowing the disturbances in δ, one can calculate the state x. We later show that
how to deal with the unknown vector δ.
Assumption A2(Input to State Stability): There exists a smooth function u(x)
such that system
x = f(x, u(x)) + δ(t) (2.11)
viewed as a system with state x and input δ(t) is input-to-state-stable.
Chapter 2. Disturbed Non-UCO Systems, the Theory 10
Considering x = f(x, u(x)) as an unforced system, this assumption implies that a
bounded disturbance δ(t) yields bounded state trajectories, so in particular it does not
drive the unforced system to instability. We will use this assumption in Section 2.4 to
prove closed-loop stability when using the observer states at the control input.
Remark 2: Assumption A2 implies that there exists a smooth function v(X) = φ(x, z)
such that the augmented system (2.7) is also input-to-state-stable. The proof is a special
case of Lemma 5.4 (ISS-Backstepping Lemma) and Corollary 5.5 in [12] when backstep-
ping and no adaptation are used. For the sake of illustration we will show a proof for a
system with two integrators (i.e. nu = 2) in the next chapter.
It is obvious that when δ(t) = 0, systems (2.11) and (2.7) are asymptotically stable with
the u(x) and v(X), respectively.
Remark 3: Notice that if in some neighborhood of the origin, [∂f/∂x] and [∂fe/∂X]
are bounded and only local properties are needed, then it will be enough to find a smooth
function u(x) such that the origin of the system x = f(x, u(x)) is asymptotically stable.
This implies that when δ(t) = 0, the origin of the extended system (2.7) is locally
stabilzing by a function of X, v(X). A proof of the local stabilizability property for (2.7)
may be found, e.g., in [10], and its global counterpart can be found in, e.g., Theorem
9.2.3 in [11] or Corollary 2.10 in [12].
When δ(t) 6= 0, once local asymptotic stability is established, by Lemma 5.4 in [14],
system (2.7) with v = φ(x, z) as the state feedback controller, viewed as a system with
state (x, z) and input δ, is locally input-to-state-stable.
Chapter 2. Disturbed Non-UCO Systems, the Theory 11
2.2 Nonlinear Observer
Assumption A2 allows us to design a stabilizing state feedback control law v = φ(x, z).
In order to perform output feedback control x should be replaced by its estimate.
Many researchers have considered undisturbed systems (δ(t) = 0) in which the knowledge
of H−11 in (2.2) is needed (e.g. [7, 8, 9]). In that case, with knowing x = H−1
1 (ye, z)
explicitly, one can estimate x = H−11 (ye, z) with the estimation of n− 1 derivatives of y,
since the vector z, containing the states of the controller, is known. Next, to estimate
the derivatives of y, a high gain observer can be employed.
Since, even if H−11 exists, it may not be possible to calculate it explicitly, we use the
observer in [16] where instead of estimating the derivatives of y and using H1(·, ·), x is
estimated directly.
The observer has the form1
˙x = f(x, z, y) = f(x, z1) +
[∂H1(x, z)
∂x
]−1
E−1L[y(t) − y(t)]
︸ ︷︷ ︸
Q
y(t) = h(x, z1)
(2.12)
where L is an n × 1 vector, E = diag [ρ, ρ2, . . . , ρn], and ρ ∈ (0, 1] is a fixed design
constant.
Notice that (2.12) does not require any knowledge of H−11 and has the advantage of
operating in x-coordinates. Assumption A1 implies that the Jacobian of the mapping
H1 with respect to x is invertible, and hence the inverse of ∂H1(x, z)/∂x in (2.12) is
well-defined.
1Throughout this section we assume A1 to hold globally, since we are interested in the ideal conver-gence properties of the state estimates. In the next section we will show how to modify the observerequation in order to achieve the same convergence properties when A1 holds over the set O ⊂ R
n ×Rnu .
Chapter 2. Disturbed Non-UCO Systems, the Theory 12
We show the properties of the observer (2.12) in in new coordinates using transformation
(2.2). We start by expressing system (2.1) in the new coordinates. By definition we have
ye4=
y
...
y(n−1)
=
h(x, u)
φ1(x, z)
...
φn−1(x, z)
+
0
θ1(δ)
...
θn−1(δ)
(2.13)
Using the definition of φn−1 in (2.6) we have
y(n) = ye,n =∂φn−1(x, z)
∂xf(x, u) +
∂φn−1(x, z)
∂xδ(t)
+nu−1∑
k=1
∂φn−1(x, z)
∂zkzk+1 +
∂φn−1(x, z)
∂znv
+∂θn−1(δ)
∂δ˙δ
Using the following definitions for α and β
(
∂φn−1(x, z)
∂xf(x, u) +
nu−1∑
k=1
∂φn−1(x, z)
∂zkzk+1
)∣∣∣∣∣(x,z)=F−1(Y,δ)
= α(ye, z)
∂φn−1(x, z)
∂znv
∣∣∣∣(x,z)=F−1(Y,δ)
= β(ye, z)v
(2.14)
we have
ye,n = α(ye, z) + β(ye, z)v +∂φn−1(x, z)
∂xδ(t) +
∂θn−1(δ)
∂δ˙δ.
Therefore, system (2.1) in the new coordinates is
ye = Acye +Bc
[
α(ye, z) + β(ye, z)v +∂φn−1(x, z)
∂xδ(t) +
∂θn−1(δ)
∂δ˙δ
]
. (2.15)
Chapter 2. Disturbed Non-UCO Systems, the Theory 13
Transforming observer (2.12) to the new coordinates
ye =[
y, ˙y, . . . , y(n−1)]>
= H1(x, z) +H2(δ) = H(x, z, δ) (2.16)
we have
˙ye,1 =∂h(x, u)
∂x˙x =
∂h(x, u)
∂xf(x, u) +
∂h(x, u)
∂xQ+
∂h(x, u)
∂uu
= φ1(x, u) + θ1(δ) − θ1(δ) +∂h(x, u)
∂xQ
= ye,2 +∂h(x, u)
∂xQ−θ1(δ)︸ ︷︷ ︸
θ1(δ)
.
For i = 2, . . . , n− 1 we have
˙ye,2 =∂φ1(x, z)
∂x˙x+
∂φ1(x, z)
∂zz +
∂φ1(x, z)
∂xQ+
∂θ1(δ)
∂δ˙δ
=∂φ1(x, z)
∂x˙x+
∂φ1(x, z)
∂zz − (θ2(δ) −
∂θ1(δ)
∂δ˙δ)
+ (θ2(δ) −∂θ1(δ)
∂δ˙δ) +
∂θ1(δ)
∂δ˙δ +
∂φ1(x, z)
∂xQ
= ye,3 +∂φ1(x, z)
∂xQ+ θ2(δ)
...
˙ye, i = ye, i+1 +∂φi−1(x, z)
∂xQ+ θi(δ)
...
where
ye, i+1 =∂φi(x, z)
∂x˙x+
∂φi(x, z)
∂zz + θi(δ)
and
θi(δ) = −(θi(δ) −∂θi−1(δ)
∂δ˙δ).
Chapter 2. Disturbed Non-UCO Systems, the Theory 14
For ˙ye,n we have
˙ye,n =∂φn−1(x, z)
∂xf(x, u) +
nu−1∑
k=1
∂φn−1(x, z)
∂zkzk+1 +
∂φn−1(x, z)
∂znu
v
+∂φn−1(x, z)
∂xQ+
∂θn−1(δ)
∂δ˙δ
=α(ye, z) + β(ye, z)v +∂φn−1(x, z)
∂xQ+
∂θn−1(δ)
∂δ˙δ
In conclusion, the observer dynamics in ye-coordinates is given by
˙ye =Acye +Bc
[
α(x, z) + β(x, z)v +∂θn−1(δ)
∂δ˙δ
]
+
[∂H1(x, z)
∂x
] [∂H1(x, z)
∂x
]−1
E−1L[y(t) − y(t)] +G(δ)
=Acye +Bc
[
α(ye, z) + β(ye, z)v +∂θn−1(δ)
∂δ˙δ
]
+ E−1L[y(t) − y(t)] +G(δ)
where (Ac, Bc) is in controllable canonical form and G(δ) is given by
G(δ) =
θ1(δ)
θ2(δ)
...
θn−1(δ)
0
.
Define ye = ye − ye as the observer error whose dynamics are
˙ye =(Ac − E−1LCc)ye
+Bc
[
α(ye, z) + β(ye, z)v − α(ye, z) − β(ye, z)v −∂φn−1(x, z)
∂xδ(t)
]
+G(δ)(2.17)
where Cc = [1, 0, · · · , 0]1×n. Equation (2.17) can be seen as a linear system with two
inputs, namely
α(ye, z) + β(ye, z)v − α(ye, z) − β(ye, z)v −∂φn−1(x, z)
∂xδ(t) (2.18)
Chapter 2. Disturbed Non-UCO Systems, the Theory 15
and G(δ).
Theorem 1 Consider system (2.8) and assume that A1 is satisfied for O = Rn+nu,
the state X belongs to a positively invariant, compact set Ω, and the following time signal
is bounded as follows (see (2.18)),
‖α(ye(t), z(t)) + β(ye(t), z(t))v(t)−α(ye(t), z(t))− β(ye(t), z(t))v(t)‖ ≤ γ1‖ye(t)− ye(t)‖
(2.19)
for some γ1 > 0, for all t ≥ 0, and for all (x(0), z(0)) ∈ Ω (a compact set) where α and
β are defined in (2.14). Choose L = [l1, · · · , ln]> such that the polynomial sn + l1sn−1 +
· · ·+ ln−1s+ ln is Hurwitz.
Under these conditions and using observer (2.12), the estimation error
‖ye‖ = ‖ye − ye‖ is uniformly ultimately bounded and its ultimate bound, which depends
on δ, can be reached arbitrarily fast.
Before proving the theorem, we clarify requirement (2.19). Requirement (2.19) is a
Lipschitz inequality which has to be satisfied at every time with a fixed Lipschitz constant.
Local Lipschitz continuity of α and β, and boundedness of x(t) and v(t) for all t ≥ 0
are not sufficient for requirement (2.19) to be satisfied. Assuming that α and β are
globally Lipschitz and v(t) is bounded will fulfill the requirement, but we will show in
a later section that using an appropriate dynamic projection for x onto a compact set,
requirement (2.19) is always satisfied and we do not need to assume global Lipschitz
continuity for α and β.
Proof. We use superposition to find the response of the observer error ye(t) for the two
inputs (2.18) and G(δ). First, we consider the case when G = 0. Using the coordinate
transformation
ν = E ′ye, E ′ 4= diag
[1
ρn−1,
1
ρn−2, . . . , 1
]
, (2.20)
Chapter 2. Disturbed Non-UCO Systems, the Theory 16
the error dynamics (2.17) is
˙ν =1
ρ(Ac − LCc)ν +Bc
[
α(ye, z) + β(ye, z)v − α(ye, z) − β(ye, z)v −∂φn−1(x, z)
∂xδ(t)
]
.
(2.21)
By assumption, Ac − LCc is Hurwitz. Let P be the solution to the Lyapunov equation
P (Ac − LCc) + (Ac − LCc)>P = −I, (2.22)
and consider the Lyapunov function candidate Vo(ν) = ν>P ν. Calculating the time
derivative of Vo along the ν trajectories we have
Vo = − ν>ν
ρ+ 2ν>PBc
[
α(ye, z) + β(ye, z)v − α(ye, z) + β(ye, z)v −∂φn−1(x, z)
∂xδ(t)
]
.
(2.23)
Using requirement (2.19) we have the following inequality
‖α(ye(t), z(t))+β(ye(t), z(t))v(t)−α(ye(t), z(t))−β(ye(t), z(t))v(t)‖ ≤ γ1‖ye(t)‖. (2.24)
Since X ∈ Ω, a compact set, and H1 is smooth, for the last element of Vo in (2.23) we
can write
∥∥∥∥
∂φn−1(x, z)
∂x
∥∥∥∥‖δ(t)‖ ≤ γ2‖δ(t)‖ (2.25)
for some positive real number γ2. Using inequalities (2.24) and (2.25) we have
Vo ≤ −‖ν‖2
ρ+ 2‖ν>‖‖P‖(γ1‖ye‖ + γ2‖δ(t)‖). (2.26)
Vo ≤ −(
1
ρ− 2‖P‖γ1
)
︸ ︷︷ ︸
η
‖ν‖2 + 2‖ν‖‖P‖γ2d (2.27)
where in the last inequality we used the fact that ‖ye‖ ≤ ‖ν‖ (see (2.20)).
Choosing 1ρ> 2‖P‖γ1, for all 0 < ε < 1 we have
Vo ≤ −η(1 − ε)‖ν‖2 ∀ ‖ν‖2 ≥ 2‖P‖εη
γ2d. (2.28)
Chapter 2. Disturbed Non-UCO Systems, the Theory 17
The ultimate bound for ‖ν‖ (and hence for ‖ye‖) is
b(ρ) =
√
λmin(P )
λmax(P )
2‖P‖ε(
1ρ− 2‖P‖γ1
)γ2d. (2.29)
It is clear that b(ρ) can be made arbitrarily small by choosing a sufficiently small ρ.
As for the time needed for ‖ν(t)‖ (and hence ‖ye(t)‖) to reach this bound, notice that
λmin(P )‖ν‖ ≤ Vo = ν>P ν ≤ λmax(P )‖ν‖. (2.30)
Using (2.30) and (2.28), we have
Vo(t) ≤−η(1 − ε)
λmax(P )Vo(t). (2.31)
Therefore, by the comparison lemma (see [14] Lemma 2.5), Vo(t) satisfies the following
inequality,
Vo(t) ≤ Vo(0) exp
−η(1 − ε)
λmax(P )t
. (2.32)
But we know that
Vo(0) = ν(0)>P ν(0) ≤ λmax(P )‖ν(0)‖2.
So (2.32) can be written as
Vo(t) ≤ λmax(P )‖ν(0)‖2 exp
−η(1 − ε)
λmax(P )t
(2.33)
An upper bound T on the time needed for the trajectory to reach the bound b in (2.29)
can be found by the following equation,
λmin(P )b2 ≤ Vo(b) ≤ λmax(P )‖ν(0)‖2 exp
−η(1 − ε)
λmax(P )t
(2.34)
Chapter 2. Disturbed Non-UCO Systems, the Theory 18
t ≥ T =λmax(P )
η(1 − ε)lnλmax(P )‖ν(0)‖2
λmin(P )b2(2.35)
and, for all t ≥ T we have ‖ν‖ ≤ b.
Using the definitions of η in (2.27) and b in (2.29), for sufficiently small ρ we can write
T ' a1ρ+ a2ρ ln(a3ρ) (2.36)
for some constants a1, a2 and a3.
Using Mercator series(ln(1 + x) = x− 1
2x2 + 1
3x3 − . . .
), we can write
limρ→0
ρ ln(a3ρ) = 0.
Therefore limρ→0 T = 0. In other words T can be made arbitrarily small by choosing a
sufficiently small ρ.
Now we return to (2.17) and continue the superposition argument by assuming that
G(δ) 6= 0 and Bc = 0, so that ˙ye = (Ac−E−1LCc)ye +G(δ). Therefore the error dynam-
ics can be viewed as the states of a linear system w = Aw+Bu, with A Hurwitz, B = 1
and u = [u1 u2 · · · un−1 0]> = G(δ). To find the solution for this system notice that
A = Ac − E−1LCc =
−L1 1 0 · · · 0
−L2 0 1 · · · 0
......
......
...
−Ln−1 0 · · · 0 1
−Ln 0 · · · 0 0
where [L1 L2 · · · Ln] = [l1/ρ l2/ρ2 · · · ln/ρ
n]. In Laplace domain we have (sI−
Chapter 2. Disturbed Non-UCO Systems, the Theory 19
A)W (s) = U(s), where the components of W (s) are given by
W1(s) =snU1(s) + sn−1U2(s) + · · · + sUn−1
sn + L1sn−1 + · · ·+ Ln−1s+ Ln
W2(s) =(s+ L1)W1(s) − U1(s) (2.37)
Wi(s) =Li−1W1(s) + sWi−1(s) − Ui−1(s) for i = 3 · · · n (2.38)
Since the poles of W1(s) are in the left-half plane, the final value theorem gives
limt→∞
w1(t) = lims→0
sW1(s) = 0.
and, as a consequence, from (2.37), when t→ ∞ we have that
‖w2(t)‖ ≤ lim sup ‖u1‖.
For i = 3 · · ·n we have
‖wi(t)‖ ≤ lim sup ‖wi−1(t)‖ + ‖ui−1(t)‖. (2.39)
To find an upper bound for ‖wi(t)‖ in (2.39), we need to find an upper bound for
‖wi−1(t)‖. To do so, notice that from (2.38) we can write recursively the following
equations
sWi−1(s) =sLi−2W1(s) + s2Wi−2(s) − sUi−2(s)
s2Wi−2(s) =s2Li−3W1(s) + s3Wi−3(s) − s2Ui−3(s)
...
si−3W3(s) =si−3L2W1(s) + si−2W2(s) − si−3U2(s)
si−2W2(s) =si−2(s + L1)W1(s) − si−2U1(s).
Chapter 2. Disturbed Non-UCO Systems, the Theory 20
Therefore when t→ ∞ we have
‖w(i−2)2 ‖ ≤ lim sup ‖u(i−2)
1 ‖ 4= u2
‖w(i−3)3 ‖ ≤ u2 + lim sup ‖u(i−3)
2 ‖ 4= u3
...
‖w(i−1)‖ ≤ u2 + · · ·+ u(i−2) + lim sup ‖u(i−1)‖4= ui−1
‖wi‖ ≤ u2 + · · ·+ ui−1 + lim sup ‖ui−1‖4= ui
which shows that for some function g of G(δ(t)) (remember that u = G(δ(t)))
limt→∞
‖w(t)‖ ≤ lim sup ‖g(G(δ(t)))‖ 4= G. (2.40)
Moreover the transient can be made arbitrarily fast by decreasing ρ.
In conclusion, using superposition, we have that lim sup ‖ye‖ ≤ G+b(ρ) where b(ρ) is the
bound in (2.29). Since b(ρ) can be made arbitrarily small and can be reached arbitrarily
fast, we can conclude that ‖ye‖ can reach any neighborhood of G arbitrarily fast by
decreasing ρ.
2.3 Projection
Expressing inequalities (2.30) as a function of ye (ν = E ′ye), we have
λmin(P )‖ye‖2 ≤ Vo = y>e E ′PE ′ye ≤1
ρ2(n−1)λmax(P )‖ye‖2. (2.41)
Note that λmin(E ′PE ′) ≥ λmin(E ′)2λmin(P ) = λmin(P ), since λmin(E ′) = 1, and
λmax(E ′PE ′) ≤ λmax(E ′)2λmax(P ) = 1/(ρ2(n−1))λmax(P ), since λmax(E ′) = 1/ρ(n−1).
Chapter 2. Disturbed Non-UCO Systems, the Theory 21
Using inequalities (2.32) and (2.41) we will have the following result,
‖ye‖ ≤√
λmax(P )
λmin(P )
1
ρn−1‖ye(0)‖ exp
−η(1 − ε)
λmax(P )t
. (2.42)
During the initial transient, ye(t) exhibits peaking, and the size of the peak grows larger
as ρ decreases and convergence rate is made faster (peaking phenomenon). In order to
isolate the peaking phenomenon from the system states, the approach generally adopted
in several papers is to saturate the control input outside a compact set of interest to
prevent it from growing above a given threshold. This technique, however, does not
eliminate the peak in the observer estimate and, hence, cannot be used to control general
systems like the ones satisfying assumption A1, since even when the system states lie in
the observable region O ⊂ Rn×R
nu , the observer estimates may enter the unobservable
domain where (2.12) is not well defined. In other words H−11 in assumption A1 is not
defined outside the observable set O, which implies that
[
∂H1(x, z)∂x
]−1
is not defined.
It appears that in order to deal with systems that are not completely observable, one
has to eliminate peaking from the observer by forcing its estimates not to exit a pre-
specified observable compact set. A common procedure in adaptive control literature for
keeping estimates in a desired convex set is to use gradient projection (see [15]). This
idea cannot be used here mainly because ˙x is not proportional to the gradient of the
observer Lyapunov function. In other words, applying a standard gradient projection
for for x over an arbitrarily convex set does not necessarily preserve the convergence
properties of the estimate. For example there are situations where, if we use projection
directly in [x>, z>]> coordinates, one can introduce equilibrium points on the boundary
of the convex set.
To solve this problem, we use projection in Y = [ye>, z>]> coordinates. In order to relate
the projected estimates in Y coordinates to [x>, z>]> coordinates, recall form (2.16) when
Chapter 2. Disturbed Non-UCO Systems, the Theory 22
δ(t) = 0 we have that ye = H1(x, z), and hence
˙ye =∂H1
∂x˙x+
∂H1
∂zz
˙x =
[∂H1
∂x
]−1(
˙ye −∂H1
∂zz
)
. (2.43)
We will use equation (2.43) to estimate x when ye is on the boundary of a specified set
C in Y = [ye>, z>]> coordinates (see equation (2.49)). In other words we confine ye and
z in that specific set C in the mentioned coordinate. In the next sections we show the
construction and properties of the set C and demonstrate how to modify the observer
using such a projection set.
2.3.1 Projection Sets
Consider system (2.8) and assume there is no disturbance δ(t). Using assumption A2
and Remark 2 we conclude the existence of a smooth stabilizing control v(X) = φ(x, z)
which makes the origin of (2.8) an asymptotically stable equilibrium point with domain of
attraction D. By the converse Lyapunov theorem found in [13], there exists a continuously
differentiable function V defined on D satisfying, for all X ∈ D,
α1(‖X‖) ≤ V (X) ≤ α2(‖X‖) (2.44)
limX → ∂D
α1(‖X‖) = ∞ (2.45)
∂V
∂X(fe(X) + ge φ) ≤ −α3(‖X‖) (2.46)
where αi, i = 1, 2, 3 are class K functions (see [14] for a definition), and ∂D is the
boundary of the set D. Given any scalar c > 0, define
Ωc4= X ∈ R
n+nu | V ≤ c
where Ωc ⊂ D for all c > 0 and, from (2.45), Ωc → D as c → ∞. Next, we define set C
as follows.
Chapter 2. Disturbed Non-UCO Systems, the Theory 23
Assumption A3(Properties of C): Assume that there exists a constant c2 > 0 and
a set C such that
F1 (Ωc2) ⊂ C ⊂ F1 (O) , (2.47)
where Ωc2 is a level set of Lyapunov function V . Assume that C has the following
properties
(i) There exists a C1 function g : C → R such that ∂C = Y ∈ C | g(Y ) = 0, and
(∂g/∂Y )> 6= 0 on ∂C, which ∂C is the boundary of C.
(ii) C z = ye ∈ Rn | [ye>, z>]> ∈ C is convex for all z ∈ R
nu .
(iii)∂g(ye, z)∂ye
6= 0 ∀z ∈ Rnu .
(iv)⋃
z∈RnuC z is compact.
F1
F−11
Ωc2
z
x
z
ye
F1(Ωc2)
C z
F1(O)
C
OF−1
1 (C)
Figure 2.2: Projection mechanism.
Chapter 2. Disturbed Non-UCO Systems, the Theory 24
In assumption A3 it is required that C possesses some topological properties which we
clarify here: part (i) means that the boundary of C is continuously differentiable, part (ii)
means that every slice of C, C z, which is obtained by holding z constant at z, is convex,
part (iii) means that normal vectors to each slice C z do not vanish anywhere on the slice
and, finally, part (iv) means that the set C is compact in the ye direction.
Remark 4: From assumption A1 we know that Y = F1(x, z) + F2(δ)4= Y 1 + F2(δ)
in which we do not know F2(δ). Therefore we use Y 1 in the projection. To do so we
need to find set C such that whenever Y 1 ∈ C we have F−11 (Y 1) ∈ O. This requirement
is represented in condition (2.47). See Figure 2.2 for a pictorial representation of this
condition. This assumption requires that there exists a set C in Y coordinates which
contains the image of Ωc2 under F1 and is contained in the image of the observable set
O under F1. This guarantees that if estimates in Y coordinates do not leave set C then
any trajectory starting in Ωc2 remains in the observable set O.
After defining C, Ωc4= X ∈ R
n+nu | V ≤ c is chosen such that it is the greatest set
which Ωc ⊂ F−11 (C).
Remark 5: To further clarify the requirements of set C, consider the sets C1 to C4 in
Figure 2.3 and Figure 2.4 for the case ye ∈ R2, z ∈ R. While they all satisfy part (i), C1
does not satisfy part (ii) since its slices along z are not convex. C2 satisfies part (ii) but
not part (iii) because the normal vector to one of the slices has no components in the
ye direction. C3 satisfies parts (i)-(iii) but not part (iv) since the area of its slices grows
unbounded as z → ∞. C4 satisfies all the properties for C.
When O = Rn × R
nu and F1(Rn+nu) = R
n+nu , A3 is always satisfied by a sufficiently
Chapter 2. Disturbed Non-UCO Systems, the Theory 25
C1 C2
Figure 2.3: Properties of C.
C3
C4
Figure 2.4: Properties of C,cont.
large set C and any c2 > 0. In order to see that, pick any c2 and choose C to be any
cylinder Y ∈ Rn+nu | ‖ye‖ ≤ M, where M > 0, containing F1(Ωc2). C always exists
since the set F1(Ωc2) is bounded. Generally, the same holds when F1(Rn+nu) is not all of
Rn+nu and Y z 4
= ye ∈ Rn |F1(x, z) is convex for all z ∈ R
nu .
When O = X × Rnu , where X is an open set which is not all of R
n, and the the system
is globally ISS (A2 holds globally), one can choose C = D × Rnu , where D ⊂ R
n is any
convex compact set with smooth boundary contained in H1(X ) and containing the origin
of ye coordinates, H1(0, 0). The scalar c2 is then the largest scalar such that F1(Ωc2) ⊂ C
(c2 does not need to be known for design purposes). In the particular case when nu = 0
Chapter 2. Disturbed Non-UCO Systems, the Theory 26
(control input u does not affect the mapping H1) and A2 holds globally, one can choose
C = D, where D is defined above.
2.3.2 Observer Estimates Projection
Recall the coordinate transformation defined in (1) and let
yPe = H1(xP , z), yPe = yPe − ye, Y P = [yPe
>, z>]>, (2.48)
where xP is the state of the projected observer defined as
˙xP = Υ(xP , z, y)
=
[∂H1
∂xP
]−1
˙y1e |xP − Γ
Nye(Y P )
(
Nye(Y P )> ˙y1
e |xP +Nz(YP )>z
)
Nye(Y P )>ΓNye
(Y P )− ∂H1
∂zz
if Nye(Y P )> ˙y1
e |xP +Nz(YP )>z ≥ 0 and Y P ∈ ∂C
f(xP , z, y) otherwise
(2.49)
where f(xP , z, y) is defined in the observer equation (2.12), ˙y1e |xP denotes the time deriva-
tive of yPe = H1(xP , z) when xP evolves according to the observer dynamics (2.12), i.e.,
˙y1e |xP =
∂H1
∂xPf(xP , z, y) +
∂H1
∂zz,
Γ = (SE ′)−1(SE ′)−>, S = S> denotes the matrix square root of P (used for the Lyapunov
function in the proof of Theorem 1) and
Nye(Y P ) =
[∂g
∂yPe(yPe , z)
]>
, Nz(YP ) =
[∂g
∂z(yPe , z)
]>
are the ye and z components of the normal vector N(Y P ) to the boundary of C at Y P ,
i.e., N(Y P ) = [Nye(Y P )>, Nz(Y
P )>]> (the function g is defined in Assumption A3).
Notice that the dynamic projection (2.49) is well-defined since assumption A3, part (iii),
guarantees that Nyedoes not vanish.
Chapter 2. Disturbed Non-UCO Systems, the Theory 27
∂C
Ns
Y P
Figure 2.5: Geometric Interpretation of the projection condition.
See Figure 2.5 for a pictorial representation of the condition in the first case of equation
(2.49). This condition is a constraint on the inner product of the two vectors s =
[ ˙y1e |xP
> z>]> and N = [Nye
>Nz>]>. Whenever Y P is on the boundary of C and the
vector field s is pointing outside of, or is tangent to, the boundary of C, this product is
greater than or equal to zero. As we mentioned earlier we want to prevent the observer
trajectories from exiting the set F−11 (C). A simple explanation of the modification in the
observer equation is that when trajectories are on the boundary of C, ∂C, we eliminate
those components of the vector field in the direction of the normal to ∂C which force the
trajectories to exit C. This in turn keeps xP within the set F−11 (C) which is a subset of
the observable set O. When xP is not on the boundary of C, we let the trajectories flow
along the vector field.
The following theorem shows that (2.49) guarantees boundedness and preserves conver-
gence for xP .
Theorem 2 : If A3 holds and (2.49) is used:
(i) Boundedness: if (xP (0), z(0)) ∈ F−11 (C), then (xP (t), z(t)) ∈ F−1
1 (C) for all t.
If, in addition, [x(t)>, z(t)>]> ∈ Ωc2 for all t ≥ 0 and the assumptions of Theorem 1 are
satisfied, then the following property holds for the flow of the projected observer dynamics
Chapter 2. Disturbed Non-UCO Systems, the Theory 28
(2.49)
(ii) Requirement (2.19) in Theorem 1 is satisfied provided supt≥0 v(t) <∞.
Notice that part (i) of the theorem shows that if xP starts in an observable subset of O,
it remains in that set for all t. It does not give any information about the behavior of
our system states x(t) or estimation error. Remark 6 deals with the estimation error and
in Section 2.4 we discuss the behavior of x(t).
Proof. In order to prove part (i) of the theorem, it is sufficient to show that set C
is a positively invariant set for Y P trajectories under the projection (2.49). This will
guarantee that (xP (t), z(t)) = F−11 (Y P ) is contained in the set F−1
1 (C) ⊂ O for all t ≥ 0.
We prove part (i) in a new coordinate [ζ>, z>]>, with the following transformation,
ζ = SE ′ye (2.50)
Similarly, ζP = SE ′yPe , ζP = SE ′yPe . Consider the mapping
G = diag [SE ′, Inu×nu]
and define C′ as the image of the set C under the linear map G, i.e.,
C′ = G(C)4=[ζ>, z>]> ∈ R
n+nu | G−1[ζ>, z>]> ∈ C.
Notice that G has no effect on z component. Figure 2.6 shows a pictorial representation
of the sets under consideration.
Define N ′ζ(ζ, z), N
′z(ζ, z) as the ζ and z components of the normal vectors to the boundary
of C′. In order to relate N ′ζ(ζ
P , z), N ′z(ζ
P , z) to Nye(yPe , z), N
′z(ζ
P , z), recall from A3 that
the boundary of C is the set ∂C = Y ∈ Rn+nu | g(Y ) = 0 and hence the boundary of C′
is the set
∂C′ = ζ ∈ Rn | g((SE ′)−1ζ, z) = 0.
Chapter 2. Disturbed Non-UCO Systems, the Theory 29
F1(Ωc2)
C
F1(O)
C z
G
G−1
C′
z
ye
z
ζ
Figure 2.6: G Transformation.
From this definition we find the expression of N ′ζ and N ′
z as
N ′ζ(ζ
P , z) = (SE ′)−>[∂g(Y P )/∂yPe ]> = (SE ′)−>Nye(Y P ),
N ′z(ζ
P , z) = Nz(YP ).
Here again we express the projected observer in ye coordinate using coordinate transfor-
mation (2.48) as follows
˙yeP =
d
dt
H1(x
P , z)
=
[∂H1
∂xP˙xP +
∂H1
∂zz
]
=
˙y1e |xP − Γ
Nye(Y P )
(
Nye(Y P )> ˙y1
e |xP +Nz(YP )>z
)
Nye(Y P )>ΓNye
(Y P )
if Nye(Y P )> ˙y1
e |xP +Nz(YP )>z ≥ 0 and Y P ∈ ∂C
˙y1e |xP otherwise
(2.51)
Chapter 2. Disturbed Non-UCO Systems, the Theory 30
Knowing that ye = (SE ′)−1ζ from (2.50), the projection (2.49) in ζ coordinates is
˙ζP = SE ′ ˙ye
P =
SE ′ ˙y1e |xP − (SE ′)−>
Nye
(
N>ye
˙y1e |xP +N>
z z)
N>ye
ΓNye
if N>ye
˙y1e |xP +N>
z z ≥ 0 and Y P ∈ ∂C
SE ′ ˙y1e |xP otherwise
(2.52)
and then substituting N ′ζ = (SE ′)−>Nye
, N ′z = Nz, and ˙y1
e |xP = (SE ′)−1 ˙ζ |xP (since we
only use ˙y1e |xP , for simplicity we do not use the notation
˙ζ1|xP ), we have
˙ζP =
˙ζ |xP −
N ′ζ
(
N ′ζ> ˙ζ |xP +N ′
z>z)
N ′ζ>N ′
ζ
if N ′ζ> ˙ζ |xP +N ′
z>z ≥ 0 and [ζP>, z>]> ∈ ∂C′
˙ζ |xP otherwise
(2.53)
Next, we show that the boundary of the set C′ is positively invariant with respect to
(2.53). In order to do that, consider the continuously differentiable function
VC′ = g ((SE ′)−1ζ, z) and calculate its time derivative along the trajectory of (2.53) when
[ζP>, z>]> ∈ ∂C′,
VC′ = N ′ζ(ζ
P , z)>˙ζP +N ′
z(ζP , z)z (2.54)
= N ′ζ> ˙ζ |xP −
N ′ζ>N ′
ζ
(
N ′ζ> ˙ζ |xP +N ′
z>z)
N ′ζ>N ′
ζ
+N ′z z (2.55)
= 0 (2.56)
which shows that the trajectory [ζP>(t)>, z>(t))]> cannot cross the boundary of C′, which
in turn implies that [yPe>(t), z>(t)]> cannot cross ∂C and, therefore, xP (t) ∈ H−1
1 (C) for
all t.
Next, for part (ii), we want to show that if X(t) = [x>(t), z>(t)]> is contained in a
positively invariant, compact set Ω for all t ≥ 0, then inequality (2.19) holds for all t ≥ 0
Chapter 2. Disturbed Non-UCO Systems, the Theory 31
using yPe (t) from part (i) instead of ye(t), provided that v(t) is uniformly bounded. Note
that z(t) is contained in the compact set Ωz = z ∈ Rnu |X(t) ∈ Ω for all t ≥ 0. Using
part (i) of this theorem and part (iv) in assumption A3 we have
[yPe>(t), z(t)>]> ∈ C =
(⋃
z∈ΩzCz)
× Ωz, for all t ≥ 0, (2.57)
where C is a compact set.
Now, part (ii) is proved by noticing that inequality (2.19) follows directly from compact-
ness of Ω, Ωz, C, and the boundedness of v(t), and the local Lipschitz continuity of α and
β. As already mentioned, the local Lipschitz continuity of α and β alone is not sufficient
for (2.19). Part (i) of this theorem is the key feature: it proves that yPe (t) is contained in
a compact set whose size is independent of ρ. This makes it possible to establish (2.19),
where the Lipschitz constant γ1 is independent of ρ (notice that bounded yPe (t) means
bounded xP (t)).
Remark 6: Notice that in Theorem 2 we do not discuss the properties of the estimation
error xP as we did for the estimation error in Theorem 1; in other words we have to show
that the properties that we established in Theorem 1 remain valid for xP . We could not
prove this part explicitly because its difficulty is beyond the scope of this thesis. Our
major problem was that we worked on a Lyapunov-based proof. But so far this path has
failed because it relies on finding an explicit Lyapunov function which can be used to
show the boundedness of observer error dynamics in equation (2.17) when G 6= 0, and not
using a superposition-based proof. The problem lies in the transformation (2.20). When
the mentioned transformation is applied, it produces two problems; the first problem
is that it renders a Lyapunov function dependent on E ′, while the second problem is
the multiplication of G(δ(t)) by E ′. At the end these two problems provide us with an
estimation error that depends on ρ, in a way that decreasing of ρ produces an increase
in estimation error.
Chapter 2. Disturbed Non-UCO Systems, the Theory 32
But based on our simulations, we conjecture that xP has the same properties as x in
Theorem 1. Therefore for now we assume that the following statement is true
Conjecture 1: Preservation of the original convergence characteristics
Properties established by Theorem 1 remain valid for xP .
For a proof of this statement when δ(t) ≡ 0 one can refer to [16], Lemma 2.
2.4 Boundedness and Closed Loop Stability
So far we have proved that the projected state estimate, xP , is bounded. Also based on
Conjecture 1 in Remark 6 we assume that the projected estimation error, xP = xP − x,
is bounded with an ultimate bound dependent on G. Here, we study the behavior of the
closed-loop system using the projected observer, i.e. the behavior of the system
X = fe(X) + gev + δe(t) (2.58)
when v = φ(xP , z).
Theorem 3 Consider system (2.58) and the set Ωc defined in Remark 4. Suppose
Assumptions A1, A2, A3 and Conjecture 1 hold. For all c ∈ (0, c), b(ρ) > 0,
d = sup(‖δ(t)‖) and 0 < θ < 1 such that
dε = α2 α−13 (A(d+ γκ(G, b(ρ)))/θ) < c (2.59)
(where α2, α3, A and κ will be defined later) there exists a number ρ ∈ (0, 1) such that
for all ρ ∈ (0, ρ] and
∀(X(0), xP (0)) ∈(X, xP ) |X ∈ Ω
c, (xP , z) ∈ F−1
1 (C),
Chapter 2. Disturbed Non-UCO Systems, the Theory 33
the phase curves X(t) of the closed-loop system remain in Ωc, and system (2.58) is
uniformly ultimately bounded.
Proof. First we show that the needed time for every trajectory starting in a set Ωc
to reach the boundary of Ωc is finite. Since (xP (0), z(0)) ∈ F−11 (C), by Theorem 2 we
have that (xP (t), z(t)) ∈ F−11 (C) for all t ≥ 0. Let Ωz
c = z ∈ Rnu | (x, z) ∈ Ωc. When
X(t) ∈ Ωc from (2.57) we have that
[xP>(t), z(t)>]> ∈ F−11
(⋃
z∈Ωzc
Cz × Ωzc
)
, for all t ≥ 0,
is a compact set independent of ρ. Hence there exists a bounded positive real number D
independent of ρ such that for all X ∈ Ωc and all (xP (t), z(t)) ∈ F−11 (C), we have that
‖fe(X) + geφ(xp, z)‖ ≤ D. Therefore we have
‖fe(X) + geφ(xp, z) + δe(t)‖ ≤ D + d.
In other words, the maximum rate change of X(t) is D+ d. Therefore ‖X(t)−X(0)‖ ≤
(D + d)t for all X(t) ∈ Ωc. If l = dist(Ωc,Ωc) = infX1∈Ωc,X2∈Ωc
‖X1 − X2‖, then T =
l/(D + d) is a uniform lower bound for the exit time from Ωc which is independent of ρ.
Now choose T0 ∈ (0, T ). From assumption A2 we know that there exists a Lyapunov
function V (X) for system (2.58) such that for the disturbance δ(·) and some positive
definite, class K∞ functions αi(·) and some K function χ(·) the following inequalities
hold
α1(‖X‖) ≤ V (X) ≤ α2(‖X‖) (2.60)
V (X) ≤ −α3(‖X‖) + χ(‖δ(t)‖) (2.61)
Choose A = supX∈Ωc‖∂V/∂X‖. From the Conjecture 1 we know that ∀t ∈ [T0, T )
there exists ρ ∈ (0, 1) such that for all ρ ∈ (0, ρ], ‖ye(t) − yPe (t)‖ ≤ G + b(ρ), which by
Assumption A1 implies that ‖x(t) − xP (t)‖ ≤ κ(G, b(ρ)) for all t ∈ [T0, T ) and some
Chapter 2. Disturbed Non-UCO Systems, the Theory 34
smooth function κ with κ(0, 0) = 0. Choose γ to be the Lipschitz constant of φ over the
set
F−11
(⋃
z∈Ωzc
Cz × Ωzc
)
.
Then we have
‖φ(x(t), z(t)) − φ(xP (t), z(t))‖ ≤ γκ(G, b(ρ))
Using a Lypunov analysis, for all t ∈ [T0, T ) we have
V =∂V
∂X
[fe(X) + geφ(xP , z) + δe(t)
]
V =∂V
∂x
[fe(X) + geφ(xP , z) − geφ(x, z) + geφ(x, z) + δe(t)
]
V =∂V
∂x[ fe(X) + geφ(x, z)] +
∂V
∂x
[geφ(xP , z) − geφ(x, z) + δe(t)
]
V ≤− α3(‖X‖) + A(‖δ‖ + γκ(G, b)) (2.62)
V ≤− α3 α−12 (V ) + A(d+ γκ(G, b)). (2.63)
For any V ≥ dε we have that
V ≤ −1 − θ
θA(d+ γκ(G, b(ρ)))
which implies that Ωc is invariant since dε < c.
Since X(t) stays in the set Ωc for all time, using (2.63) and, e.g. Lemma 5.3 in [14], we
can find an ultimate bound for X(t) which is given by
M = α−11
(
α2
(
α−13
(A(d+ γκ(G, b(ρ))
θ
)))
(2.64)
The main result of this theorem is to show that if we use the projected observer in the
the feedback loop of the extended system (2.58), or in other words v = φ(xP , z), with
a sufficiently small ρ, we can guarantee that the observer estimates will converge to a
neighborhood of the system states in finite time and before the system states exit the
Chapter 2. Disturbed Non-UCO Systems, the Theory 35
set Ωc. When ρ is made smaller, the projected observer does not allow large peaks in
the observer estimates, therefore preventing large control inputs which may drive X(t)
outside the set Ωc in shorter time.
When X(t) stays in Ωc, the Lyapunov analysis shows that X(t) is ultimately bounded.
With (2.59) we guarantee that this bound is inside the set Ωc. The major difficulty here
is to find an analytical solution for c and d (d = sup ‖δ(t)‖) such that (2.59) holds. Notice
that from (2.29) and (2.40), when δ(t) = 0 (d = 0), we have that M = 0, which implies
that X(t) approaches the origin. For a complete discussion on asymptotic stability of
the origin when δ(t) = 0 and the projected observer is used, one can refer to [16].
Notice that in general, when the extended system X = fe(X) + geφ(xP , z) is asymp-
totically stable at the origin, adding small disturbance δe(t) to the system (see (2.58))
will render the extended system UUB. In this way, the domain of attraction is local and
the bound for δ(t) is not known, while in our method we have a regional result and a
bound for the disturbance defined in (2.59).
2.5 Concluding Remarks
In this chapter we developed an output feedback control theory for a class of nonlinear sys-
tems subjected to disturbances. We specified allowable disturbances and system models
by imposing some restrictions on the observability mapping (2.2). Next, by introducing
the projected observer (2.49) in the feedback loop and the concept of projection sets,
we showed that we could achieve boundedness and stability in the presence of bounded
disturbances.
One of the benefits of the approach introduced in this chapter is that the knowledge of the
inverse of the observability mapping H1 is not needed. Notice that the ISS assumption
A2 does not affect our assumptions regarding observability mapping or disturbances,
Chapter 2. Disturbed Non-UCO Systems, the Theory 36
which implies that, having observability assumption A1 satisfied for a system, we can
use standard state feedback control design techniques to achieve input to state stability.
Although, when the disturbance δ(t) is small, one can get local results by the method
introduced in this chapter for a linearized model of the original system (2.1), notice that
the observer designed in this way would not preserve the boundedness for the original
nonlinear system.
In the next chapters we will describe the difficulties that may be encountered when using
this technique in practice.
Chapter 3
Application to Axial Flow
Compressors
In this chapter we apply the techniques we developed in the previous chapter to the surge
and stall problem of axial air compressors. Surge and stall are instability phenomena
limiting the range of operation for compressors.
The instability problem has been studied and industrial solutions based on surge avoid-
ance are well established. These solutions are based on keeping the operating point to
the right of the compressor surge line using a surge margin (this point will be explained
shortly). However, there is potential for increasing the efficiency of compressors by allow-
ing the operation closer to the surge line and increasing the range of mass flow over which
the compressor can operate stably. This, however, raises the need for control techniques
which stabilize the compressor also to the left of the surge line. This approach is known
as active surge control.
In the following sections, we give a brief introduction to compressors, axial flow com-
pressor model, and surge and stall problem. Next we use the technique in Chapter 2
to stabilize axial flow compressors in the presence of disturbances. The introduction to
compressors, surge and stall, and Moore-Greizter model are based on the material in [31],
37
Chapter 3. Application to Axial Flow Compressors 38
[38] and [39]. Parts of the state feedback stabilizer design and the projection set rely on
the work in [23].
3.1 Compressors
3.1.1 Types of Compressors
The function of a compressor can be defined as to raise pressure of a specified mass flow
of gas by a prescribed amount using the minimum power input.
Compressors are used in a variety of applications such as, turbojet engines used in
aerospace propulsion, power generation using industrial gas turbines, pressurization of
gas and fluids in the process industry, transport of fluids in pipelines and so on.
Compressors can be divided into four general types: reciprocating, rotatory, centrifugal
and axial. Reciprocating, rotatory and centrifugal compressors are not studied in this
work. Axial compressors, also known as turbo-compressors or continuous flow compres-
sors, work by the principle of accelerating the fluid to a high velocity and then converting
the kinetic energy and pressure rise by decelerating the gas in diverging channels. In axial
compressors the deceleration takes place in the stator blade passages.
3.1.2 The Axial Compressor, Principles of Operation
A stage of an axial compressor consists of a row of rotor blades and a row of stator blades,
as shown in Figure 3.1, where C1, C2 are the velocities of the flow through the rotor and
stator respectively, C3 is the velocity of the flow exiting the stator, and U is the rotor
velocity.
A single stage has often a low pressure ratio 1.1:1 to 1.2:1. Therefore it is very common
to use multi-stage axial compressors whereby the compressor has a series of stages, each
Chapter 3. Application to Axial Flow Compressors 39
having a row of rotor blades and a row of stator blades, where the number of stages
depends of the desired pressure ratio.
Rotor
Stator
C3
C2
C1
U
Figure 3.1: Blade rows.
In axial compressors, gas is accelerated in the rotor blades, and decelerated in the stator
blades, converting kinetic energy gained in the rotor into pressure rise. All the power
is transferred to the gas in the rotor, the stator merely transforms kinetic energy to an
increase in static pressure with the temperature remaining constant.
3.1.3 Surge
Surge is an axisymmetrical oscillation of the flow through the compressor, and is char-
acterized by a limit cycle in the compressor characteristic. Surge oscillations are in most
applications unwanted, and in extreme cases can damage the compressor.
The mechanism behind surge can be explained as follows. First, consider a compressor
operating between two constant pressure reservoirs. The compressor is equipped with a
downstream throttle valve. First the compressor operates at point A on the compressor
Chapter 3. Application to Axial Flow Compressors 40
characteristic of Figure 3.2. Consider a disturbance, in the form of partially closing the
Pressure
Mass Flow
C
B A
Figure 3.2: Compressor characteristic.
throttle valve and thereby temporarily reducing the flow (not beyond the maximum of the
characteristic). This results in an increase in the delivery pressure from the compressor
and a reduction in the compressor flow. The increased delivery pressure produces a
larger mass flow trough the throttle, reducing compressor delivery pressure and increasing
compressor flow. This is therefore self compensating, stable system.
Now, consider the compressor operating at point B in Figure 3.2. A reduction in the mass
flow would now result in reduced compressor delivery pressure, reducing flow through the
throttle, moving the operating point further and further to the left. Eventually, the mass
flow reduction would be so great that the pressure upstream of the throttle falls below
the compressor delivery pressure. Mass flow will then increase until the system is drawn
back to the operating point B, and the whole cycle repeats. This instability phenomenon
is referred to as surge.
As a rule, stall and surge occur at the local maximum of the compressor characteristic or
at a point of the compressor characteristic with a certain positive slope. The surge point
will be located at some small distance to the left of the peak. Figure 3.3 is a representation
of the compressor characteristic for different values of rotor speed (ni’s). Surge points are
located on the so-called surge line, which can be considered as a barrier that separates
Chapter 3. Application to Axial Flow Compressors 41
Pressure
Mass Flow
n1
n2
n3
n4
speed
surge line
surge avoidance line
Figure 3.3: Compressor characteristic with surge and surge avoidance lines.
regions of stable and unstable operation in the compressor map. In surge avoidance
techniques, a surge avoidance line in the compressor map is introduced which is in some
distance (e.g., 10% of the flow rate) from the actual surge line, and the compressor state
is not allowed to cross to the left of this line, see Figure 3.3.
3.1.4 Rotating Stall
Rotating stall is an instability which arises when the circumferential flow pattern is
disturbed. Rotating stall occurs when one or more stall cells of reduced, or stalled flow
propagate around the compressor annulus at the fraction of the rotor speed. According
to Greitzer, the stall cell propagation may be 20-70% of the rotor speed. Rotating stall
leads to a reduction of the pressure rise of the compressor, which in the compressor
characteristic corresponds to the compressor operating on the in-stall characteristic1, see
Figure 3.4 (this point will be made clear in what follows).
A basic explanation of the rotating stall mechanism can be summarized as follows. Con-
sider a row of axial compressor blades operating at a high angle of attack, as shown in
1In-stall or stalled flow characteristic is a set of operating points which define the pressure rise andmass flow relation when the compressor is in stall. These operating points may be stable and thecompressor remain there, until measures are taken to bring it back to the un-stalled characteristic. InFigure 3.4 un-stalled characteristic is indicated by compressor char. A technique for extracting thein-stall characteristic for an axial flow compressor is presented in [22].
Chapter 3. Application to Axial Flow Compressors 42
(1)
(3)
(4)
pressure
mass flow
compressorchar.In-stall char.
(2)
Figure 3.4: Shematic drawing of hysteresis caused by rotating stall. Solid lines represent
stable equilibria and dotted lines represent unstable equilibria. The dashed lines are the
throttle lines (throttle pressure-flow characteristic) for the onset and clearing of stall.
Figure 3.5. Suppose that there is a non-uniformity in the inlet flow such that a locally
higher angle of attack is produced on blade B which is enough to stall it. The flow now
separates from the suction surface of the blade, producing a flow blockage between B
and C. This blockage causes a diversion of the inlet flow away from B towards A and
C, resulting in an increased angle of attack on C, causing it to stall. Thus the stall cell
propagates along the blade row.
Another consequence of rotating stall is hysteresis occurring when trying to clear the
stall by using the throttle. This situation is shown in Figure 3.4 and can be described
as follows. Initially the compressor is operating stably (1), then a disturbance drives
the compressor to rotating stall, and an operating point on the low pressure in-stall
characteristic (2). By opening the throttle to clear the stall, a mass flow corresponding
to (4) which is higher than the initial mass flow corresponding to (1), is required, before
the operating point is back on the un-stalled compressor characteristic.
Chapter 3. Application to Axial Flow Compressors 43
Direction of
stall propagationC
B
A
Figure 3.5: Physical mechanism for inception of rotating stall.
3.1.5 Modelling of Axial Compression Systems
Although dynamic models of basic compression systems have been available since 1955, a
major step in this field was made in 1976 when a nonlinear dynamic model for a basic axial
compression system was presented by Greitzer. A major drawback of earlier models was
that they were linearized, and therefore unable to describe the large amplitude pulsations
during the surge cycle due to their restriction to small perturbations from an equilibrium.
The main contribution of Greitzer is the finding of B parameter (which will be defined
shortly) and showing that B > Bcrit leads to surge and B < Bcrit leads to rotating stall.
While the model of Greitzer is capable of simulating surge oscillations, rotating stall
is described as a pressure drop. Motivated by stall problems in gas turbines, Moore
and Greitzer (1986) proposed a model for multistage axial compressors where rotating
stall is included as a state. The low order model of Moore and Greitzer captures the
Chapter 3. Application to Axial Flow Compressors 44
post stall transients of a low speed axial compressor-plenum-throttle model (as shown in
Figure 3.6).
Lc
LE
Vp
Inlet Duct
Compressor
Exit Duct
Plenum
Throttle
LI
Figure 3.6: Compressor system.
The three differential equations of the model resulting from a Galerkin approximation of
the original PDE model are
Ψ1 =W/H
4B2
(Φ1
W− 1
WΦT1(Ψ1)
)H
Lc(3.1)
Φ1 =H
Lc
(
−Ψ1 − ψ1,c0
H− 1
2
(Φ1
W− 1
)3
+ 1 +3
2
(Φ1
W− 1
)(
1 − J
2
))
(3.2)
J = J
(
1 −(
Φ1
W
)2
− J
4
)
σ (3.3)
where
• H is the semi-height of the compressor characteristic (see Figure 3.7)
• W is the semi-length of the compressor characteristic (see Figure 3.7)
Chapter 3. Application to Axial Flow Compressors 45
• Φ1 is the compressor mass flow
• Ψ1 is the plenum pressure rise
• J is the squared amplitude of the rotating stall
• ΦT1(Ψ1) is the mass flow through the throttle
• Lc is the effective flow-passage length of the compressor and the ducts defined as
Lc4= LI +
1
a+ lE
where the positive constant a is the reciprocal time-lag parameter of the blade
passage.
• Constant B > 0 is the Greitzer B-parameter defined as
B4=
U
2as
√
VpAcLc
where U is the constant compressor tangential speed, as is the speed of sound, Vp
is the plenum volume, Ac is the flow area and Lc is the length of the duct and the
compressor.
• Constant σ > 0 is defined as
σ =3aH
(1 +ma)W
where m is the compressor-duct flow parameter.
Notice that all the time derivatives in equations (3.1)-(3.3) are calculated with respect
to the non-dimensional time ξ1, defined as
ξ14=Ut
R
where t is the actual time and R is the mean compressor radius.
Chapter 3. Application to Axial Flow Compressors 46
The pressure rise of the compressor is a nonlinear function of the mass flow. This function,
Ψ1,c(Φ1), is known as the compressor characteristic. One expression for this characteristic
which has found widespread acceptance in the control literature is the characteristic of
Moore and Greitzer
Ψ1,c(Φ) = ψ1,c0 +H
(
1 +3
2
(Φ1
W− 1
)
− 1
2
(Φ1
W− 1
)3)
(3.4)
The characteristic is shown in Figure 3.7. Notice that in (3.2)
Φ1 =H
lc
(−Ψ1
H+
Ψ1,c(Φ1)
H− 3
4J
(Φ1
W− 1
))
. (3.5)
Ψ
Φ
ψc0
Ψc(Φ)
2W
2H
Figure 3.7: Cubic compressor characteristic of Moore and Greitzer. The constants W
and H are known as the semi width and semi hight, respectively.
3.1.6 Disturbances
As in [19], the effect of pressure and flow disturbances are considered in this chapter.
Pressure disturbances, which may arise from combustion-induced fluctuations, accelerate
the flow. Flow disturbances may arise from processes upstream of the compressor, other
Chapter 3. Application to Axial Flow Compressors 47
compressors in series, or an air cleaner in the compressor in-let duct (see Figure 3.6). In
case of an aircraft jet engine, large angle of attack or altitude variations may cause mass
flow disturbances.
Here we assume that disturbances, δ(t), are bounded, and they are either slow-varying or
constant. The constant disturbance case is of particular interest when, e.g., a constant
negative mass flow disturbance pushes the equilibrium over the surge line, initiating
surge or rotating stall. A constant disturbance in equation (3.1) can be considered as an
uncertainty in the throttle characteristic. Likewise a constant disturbance in equation
(3.2) can be considered as an uncertainty in pressure characteristic Ψ1,c.
3.1.7 Change of Variables
Consider system (3.1)-(3.3). We use the following change of variables to obtain a different
form for the Moore-Greitzer equations which is easier to use. The change of variables is,
Φ =Φ1
W− 1 Ψ =
Ψ1
WR =
J
4ξ =
H
WLcξ1. (3.6)
Applying the above changes, system (3.1)-(3.3) is given by
Φ = −Ψ + ΨC(Φ) − 3ΦR
Ψ =1
β2(Φ − ΦT )
R = σR(1 − Φ2 − R), R(0) ≥ 0
(3.7)
where β = 2BHW
, Φ represents the mass flow, Ψ is the plenum pressure rise, R ≥ 0 is the
normalized stall cell squared amplitude, and ΦT is the mass flow through the throttle.
Referring to (3.4) here we have ΨC(Φ) = Ψ1,c
H= ΨC0 +1+ 3
2Φ− 1
2Φ3 where ΨC0 = ψ1,c0
H.
ΦT (Ψ) is the throttle characteristic, and is defined as Ψ = 1γ2 (1 + ΦT (Ψ))2, where γ is
the throttle opening, the control input. For our purpose we assign σ = 7, and β = 1/√
2.
We call (3.7) the MG3 (Moore-Greitzer three state) model.
Chapter 3. Application to Axial Flow Compressors 48
3.2 Output Feedback Control for MG3
Our control objective is to stabilize system (3.7) around the critical equilibrium Re =
0,Φe = 1,Ψe = ΨC(Φe) = ΨC0+ 2, which achieves the peak operation on the compressor
characteristic, in the presence of external disturbances. We shift the origin to the desired
equilibrium with the change of variables φ = Φ− 1, ψ = Ψ−ΨC0− 2. System (3.7) then
becomes
R = −σR2 − σR(2φ+ φ2)4= f1(R, φ, ψ)
φ = −ψ − 3
2φ2 − 1
2φ3 − 3Rφ− 3R
4= f2(R, φ, ψ)
ψ = − 1
β2(ΦT − 1 − φ).
(3.8)
We assume the pressure rise (and hence ψ) to be the only measurable state variable.
System (3.8) including disturbances can be represented as following
x1 = f1(x) + δ1(t) δ1(t) > 0
x1 = f2(x) + δ1(t)
x3 = kx2 − ku+ δ3(t)
y = x3
(3.9)
where [x1, x2, x3] = [R, φ, ψ], u = ΦT − 1, k = − 1β2 , and δi is disturbance.
It is clear that the system is in the form of system (2.1).
3.2.1 Observability Mappings
First we have to check the observability mapping in Assumption A1 of Chapter 2 to
see if F has the desired form. The observability mapping for system (3.9) including
disturbances is
Chapter 3. Application to Axial Flow Compressors 49
ye1 = y = x3
ye2 = y = x3 = kx2 − ku+ δ3(t)
ye3 = y = kx2 − ku+ δ(t) = kf2(x) − ku+ kδ2(t) + δ3(t),
(3.10)
so letting z = [u, u]> and δ = [δ, δ]>, H1 and H2 in Assumption A1 have the expression
H(x, z, δ) = H1(x, z) +H2(δ)
=
x3
kx2 − ku
kf2(x) − ku
+
0
δ3(t)
kδ2(t) + δ3(t)
=
h(x)
φ1(x, z)
φ2(x, z)
+
0
θ1(δ)
θ2(δ).
(3.11)
Notice that ∂H∂x
= ∂H1∂x
. Therefore F in A1 can be written as
F (R, φ, ψ, z1, z2) =
ψ
1/β2 (φ− z1))
1/β2 (−ψ − 3/2φ2 − 1/2φ3 − 3Rφ− 3R− z2)
z1
z2
+
0
δ3(t)
kδ2(t) + δ3(t)
0
0
= F1(x, z) + F2(δ).
(3.12)
It is clear that F has the desired form. Notice that when φ = −1, which corresponds to
Φ = 0, F1 does not depend on R and is not invertible. Therefore when there is no mass
flow through the compressor (Φ = 0), a condition that we want to avoid during normal
Chapter 3. Application to Axial Flow Compressors 50
operation, R cannot be determined. In all other cases F1 is a deffeomorphism. We thus
conclude that assumption A1 is satisfied on the set
O =[R, φ, ψ]> ∈ R
3, z ∈ R2 |φ > −1
.
3.2.2 State Feedback Control
Here we design a state feedback control law for system (3.8) and later we show that
this control law input-to-state stabilizes system (3.8). By viewing system (3.8) as an
interconnection of two subsystems, namely the R-subsystem and the (φ, ψ)-subsystem,
we design a full-state feedback controller which makes the origin of (3.8) an asymptotically
stable equilibrium point with domain of attraction (R, φ, ψ) ∈ R3|R ≥ 0. We begin by
assuming that δ(t) = 0.
The following theorem and its proof are taken from [23] and will be used in the following
to prove the ISS property in assumption A2 from Chapter 2.
Theorem 4 For system (3.8), with the choice of the control law
u = (1 − β2k1k2)φ+ β2k2ψ + 3β2k1Rφ (3.13)
where k1 and k2 are positive scalars satisfying the inequalities,
k1 >17
8+
(2Cσ + 3)2
2(3.14)
(
Cσ − 105
64
)
k21 +
3
4
(
−1
2Cσ +
21
4
)
k1 − (Cσ + 3)2 > 0 (3.15)
k2 > k1 +9
4k2
1 +9k1
4k1 − 9/2+
(k21 − 1)2
4(3.16)
C >3
2σ(3.17)
the origin is an asymptotically stable equilibrium point with domain of attraction A =
(R, φ, ψ) ∈ R3|R ≥ 0.
Chapter 3. Application to Axial Flow Compressors 51
Proof. Without loss of generality and for simplicity let
u′ =1
β2(u− φ),
be the control input. Therefore the last equation in (3.8) becomes ψ = −u′. In the next
step, we consider system (3.8) the interconnection of two subsystems [S1] and [S2] as
follows
[S1] R = −σR2, [S2]
φ = −ψ − 3
2φ2 − 1
2φ3
ψ = −u′
A Lyapunov function for [S1], defined on the domain R ∈ R |R ≥ 0, is V1 = R, and its
time derivative is V1 = −σR2 thus showing that the origin of [S1] is an asymptotically
stable equilibrium point of [S1], and its domain of attraction is R ∈ R |R ≥ 0. As for
subsystem [S2] we use V2 = 12φ2+ k1
8φ4+ 1
2(φ−k1ψ)2, where k1 is a positive design constant.
Furthermore, in [12], a stabilizing control law for [S2] is found to be u = −c1φ + c2ψ,
where c1 and c2 are two appropriate positive constants. In the following we will show
that, in order to stabilize the interconnection of systems [S1] and [S2], one needs to add
to u = −c1φ + c2ψ a term which is proportional to the product Rφ. Based on these
considerations, consider the following candidate Lyapunov function for system (3.8),
V = CV1 + V2 = CR+1
2φ2 +
k1
8φ4 +
1
2(ψ − k1φ)2 (3.18)
where C > 0 is a scalar. Notice that V is positive definite on the domain A. Letting
ψ = ψ − k1φ, we calculate the time derivative of V as follows,
V = − CσR2 − CσR(2φ+ φ2) +
(
φ+k1
2φ3
)(
−ψ − 3
2φ2 − 1
2φ3 − 3Rφ− 3R
)
+
+ ψ
(
−u′ + k1ψ +3
2k1φ
2 +1
2k1φ
3 + 3k1Rφ+ 3k1R
)
(3.19)
We use the identity −32φ2 − 1
2φ3 = −1
2
(φ+ 3
2
)2φ+ 9
8φ to eliminate the potentially desta-
bilizing term − (φ+ k1/2φ3) 3/2φ2. Next, substituting (3.13) into (3.19) (after taking
Chapter 3. Application to Axial Flow Compressors 52
into account the definition of u′), letting k1 = k1 − 9/8, we get
V = − CσR2 − CσR(2φ+ φ2) +
(
φ+k1
2φ3
)(
−ψ − k1φ− 1
2
(
φ+3
2
)2
φ− 3Rφ− 3R
)
+
+ ψ
(
−(k2 − k1)ψ + k21φ+
3
2k1φ
2 +1
2k1φ
3 + 3k1R
)
(3.20)
Notice that the expression −φ
2
(φ+ k1
2φ3) (φ+ 3
2
)2can be discarded since it is negative
definite, and that the term k12φ3ψ cancels out. After rearranging the remaining terms,
we get
V ≤− CσR2 − (2Cσ + 3)Rφ− (Cσ + 3)Rφ2 − k1φ2 −
(k1k1
2+
3k1
2R
)
φ4 − 3k1
2Rφ3+
+ ψ
(
−(k2 − k1)ψ + (k21 − 1)φ+
3
2k1φ
2 + 3k1R
)
(3.21)
By using Young’s inequality2 we obtain the following inequalities
− (2Cσ + 3)Rφ ≤ 1
2R2 +
(2Cσ + 3)2
2φ2, −3k1
2Rφ3 ≤ 3k1
2
(Rφ2
4+Rφ4
)
,
(k21 − 1)φψ ≤ φ2 +
(k21 − 1)2
4ψ2, 3k1Rψ ≤ R2 +
9
4k2
1ψ2,
3
2k1φ
2ψ ≤ k1k1
4φ4 +
9k1
4k1
ψ2.
(3.22)
Applying the above inequalities to (3.21) we get
V ≤−(
Cσ − 3
2
)
R2 −(
k1 −(2Cσ + 3)2
2− 1
)
φ2 −(
k2 − k1 −9
4k2
1 −9k1
4k1
− (k21 − 1)2
4
)
ψ2+
−(
Cσ + 3 − 3
8k1
)
Rφ2 − k1k1
4φ4,
≤−
R
φ2
>
Cσ − 32
12
(Cσ + 3 − 3
8k1
)
12
(Cσ + 3 − 3
8k1
)14k1k1
R
φ2
−
(
k1 −(2Cσ + 3)2
2− 1
)
φ2+
−(
k2 − k1 −9
4k2
1 −9k1
4k1
− (k21 − 1)2
4
)
ψ2 (3.23)
If the above quadratic form is positive definite, and the coefficients multiplying φ2 and
ψ2 are positive, then V is negative definite on the domain A. For the quadratic form to
2For any real numbers a and b, and any positive real k, one has that ab ≤ a2
4k+ kb2.
Chapter 3. Application to Axial Flow Compressors 53
be positive definite we should have
Cσ − 3
2> 0 ,
(
Cσ − 3
2
)1
4k1k1 −
1
4
(
Cσ + 3 − 3
8k1
)2
> 0.
For positivity of the coefficients of φ2 and ψ2 we need that
k1 >(2Cσ + 3)2
2+ 1 , k2 > k1 +
9
4k2
1 +9k1
4k1
+(k2
1 − 1)2
4.
By using the definition of k1, inequalities (3.25), (3.26), (3.27), and (3.28) follow. In con-
clusion, if k1, k2, and C are chosen so that (3.25)-(3.28) hold, V will be negative definite
on A which contains the origin. Moreover, the boundary of A, ∂A = (R, φ, ψ) |R = 0,
is an invariant manifold (when R = 0, R = 0). Therefore the origin of the closed-loop
system in an asymptotically stable equilibrium point and the set (R, φ, ψ) | V ≤ K∩A
is its region of attraction for any positive real number K. This in turn shows that A is
the domain of attraction of the origin of the closed-loop system.
3.2.3 Input-to-State Stability
Now we have to check assumption A2 for system the (3.9) using the feedback controller
u in (3.13).
Consider
R = f1(x) + δ1(t)
φ = f2(x) + δ2(t)
ψ = − 1
β2(u− φ) + δ3(t)
(3.24)
where f1 and f2 are defined in (3.8), δi’s are disturbances, and x = [R, φ, ψ]>.
Chapter 3. Application to Axial Flow Compressors 54
Theorem 5 The closed-loop system given by (3.24) and (3.13) where k1 and k2 satisfy
inequalities
k1 >17
8+
(2Cσ + 3)2
2(3.25)
(
Cσ − 105
64
)
k21 +
3
4
(
−1
2Cσ +
21
4
)
k1 − (Cσ + 3)2 > 0 (3.26)
k2 > k1 +9
4k2
1 +9k1
4k1 − 9/2+
(k21 − 1)2
4(3.27)
C >3
2σ(3.28)
is input-to-state stable with respect to the disturbance input [δ1(t), δ2(t), δ3(t)]>.
Proof. Substituting u = u from (3.13) in the last equation of (3.24) we have the following
u = (1 − β2k1k2)φ+ β2k2ψ + 3β2k1Rφ4= (1 − a)φ+ bψ + dRφ
ψ = − 1
β2(−aφ + bψ + dRφ)
︸ ︷︷ ︸
f3
+δ3(t)
ψ = f3(x) + δ3(t).
We have the following closed loop system
R = f1 + δ1
φ = f2 + δ2
ψ = f3 + δ3
(3.29)
where we assume that δi’s are bounded.
Using the same Lyapunov function V = CV1 + V2 (C > 0), as in the state feedback
controller design section, we have the following
V1 = R, V2 =1
2φ2 +
k1
8φ4 +
1
2(ψ − k1φ)2
︸ ︷︷ ︸
ψ
Chapter 3. Application to Axial Flow Compressors 55
V = CR +∂V2
∂φφ+
∂V2
∂ψ
˙ψ
= Cf1 + Cδ1 +∂V2
∂φf2 +
∂V2
∂φδ2 + ψ(ψ − k1φ)
= Cf1 +∂V2
∂φf2 + ψ(f3 − k1f2)
︸ ︷︷ ︸
M
+Cδ1 +∂V2
∂φδ2 + ψδ3 − k1ψδ2
︸ ︷︷ ︸
N
.
Using Young’s inequality for N we will have
N = Cδ1 + (φ+k1
2φ3)δ2 + ψδ3 − k1ψδ2
≤ Cδ1 +φ2
2+δ22
2+k1
8φ6 +
k1
2δ22 +
ψ2
2+δ23
2+k1
2ψ2 +
k1
2δ22
≤ k1
8φ6 + Cδ1 +
φ2
2+δ22
2+k1
2δ22 +
ψ2
2+δ23
2+k1
2ψ2 +
k1
2δ22
︸ ︷︷ ︸
M ′
In the controller design section we discarded the negative definite term −φ
2
(φ+ k1
2φ3) (φ+ 3
2
)2,
which is used here as follows
−φ2
(φ+k1
2φ3)(φ+
3
2)2 = −φ
2
2(φ+
3
2)2
︸ ︷︷ ︸
negative definite
−k1
4φ4(φ+
3
2)2
≤ −k1
4φ6 − 3
4k1φ
5 − 9
16k1φ
4
then for V we have
V ≤M − k1
4φ6 − 3
4k1φ
5 − 9
16k1φ
4 +k1
8φ6 +M ′
≤M − k1
4φ6 − 3
4k1φ
5 − 9
16k1φ
4 +M ′
≤M − k1
2φ4(
φ2
4+
3
2k1φ+
9
8) +M ′
≤M − k1
2φ4
[
(φ
2+
3
2)2 − 9
8
]
+M ′
≤M −k1
2φ4(
φ
2+
3
2)2
︸ ︷︷ ︸
negative definite
+9
16k1φ
4 +M ′
Chapter 3. Application to Axial Flow Compressors 56
Hence V satisfies the following inequality,
V ≤−(
Cσ − 3
2
)
R2 −(
k1 −(2Cσ + 3)2
2− 1
)
φ2
−(
k2 −3
2k1 −
9
4k2
1 − 9k1
k1
− (k1 − 1)2
4− 1
2
)
ψ2
− (Cσ + 3 − 3
8k1)Rφ
2 −(k1k1
4− 9k1
16
)
φ4
+ Cδ1 +
(1
2+ k1
)
δ22 +
δ23
2
≤−
R
φ2
T
Cσ − 32
12(Cσ + 3 − 3
8k1)
12(Cσ + 3 − 3
8k1)
k1k1
4− 9k1
16
R
φ2
−(
k1 −(2Cσ + 3)2
2− 1
)
φ2 −(
k2 −3
2k1 −
9
4k2
1 − 9k1
k1
− (k1 − 1)2
4− 1
2
)
ψ2
+ Cδ1 +
(1
2+ k1
)
δ22 +
δ23
2
(3.30)
Note that
Cδ1 +
(1
2+ k1
)
δ22 +
δ23
2≤ C|δ1| +
(1
2+ k1
)
(δ22 + δ2
3)
≤ C‖δ‖ +
(1
2+ k1
)
‖δ‖2 4= χ(||δ||)
If the parameters of the inequality (3.30) satisfy the following inequalities,
k1 ≥17
8+
(2Cσ + 3)2
2(3.31)
(
Cσ − 105
64
)
k21 +
3
4
(
−5
4Cσ +
51
8
)
k1 − (Cσ + 3)2 ≥ 0 (3.32)
k2 ≥3
2k1 −
9
4k2
1 − 9k1
4k1 − 9\2 − (k1 − 1)2
4− 1
2(3.33)
Cσ ≥ 105
64(3.34)
Chapter 3. Application to Axial Flow Compressors 57
then the quadratic form in equation (3.30) is positive definite and the coefficients of φ
and ψ are positive and we have
V ≤−
R
φ2
T
Cσ − 32
12(Cσ + 3 − 3
8k1)
12(Cσ + 3 − 3
8k1)
k1k1
4− 9k1
16
R
φ2
−(
k1 −(2Cσ + 3)2
2− 1
)
φ2 −(
k2 −3
2k1 −
9
4k2
1 − 9k1
k1
− (k1 − 1)2
4− 1
2
)
ψ2
+ χ(||δ||)4= −α(‖x‖) + χ(||δ||)
which implies that the system (3.29) is input-to-state stable. Also notice that the set of
triples (k1, k2, σ) satisfying (3.31)-(3.34) is non-empty and they also satisfy inequalities
(3.25)-(3.28)
3.2.4 Stabilizing Control Law for Augmented System
Referring to equations (3.10), since H1 depends on u and u (i.e. nu = 2), following the
procedure outlined in Section 2.1 we augment the system dynamics with two integrators
at the input side, i.e.,
[P1]
R = −σR2 − σR(2φ+ φ2) + δ1(t)
φ = −ψ − 3/2φ2 − 1/2φ3 − 3Rφ− 3R + δ2(t)
ψ = 1β2 (φ− z1) + δ3(t)
[P2]
z1 = z2
z2 = v
(3.35)
Defining x = [R, φ, ψ]>, we can write system [P1] as x = f(x) + g(x)z1 + δ(t). The
following theorem shows that having the stabilizing controller (3.13), we can find a v
such that the extended system is input-to-state-stable.
Chapter 3. Application to Axial Flow Compressors 58
Theorem 6 Consider system
x = f(x) + g(x)u+ δ(t) (3.36)
Suppose there exists a smooth function u = u0(x) such that system (3.36) is ISS with δ(t)
as an input. Then there exists a smooth function v such that system
x = f(x) + g(x)s1 + δ(t)
s1 = s2
s2 = v
(3.37)
is ISS with δ(t) as the input.
Proof. Consider the following change of variables
z1 = s1 − u0(x) (3.38)
z2 = s2 − u1(x, s1) (3.39)
Then we have
x = f(x) + g(x)s1 + δ(t)
z1 = z2 + u1 −∂u0
∂x[f + g(z1 + u0)]
︸ ︷︷ ︸
A
−∂u0
∂xδ
z2 = v − ∂u1
∂x[f + g(z1 + u0)]
︸ ︷︷ ︸
B
−∂u1
∂xδ − ∂u1
∂s1(z2 + u1)
(3.40)
From ISS property for the system (3.36) we know that there exists a Lyapunov function
V (x) for that system such that for the disturbance δ(·) and some positive definite, class
K∞ functions αi(·) and some K function χ(·) the following inequalities hold
α1(‖x‖) ≤ V (x) ≤ α2(‖x‖) (3.41)
V (x) ≤ −α3(‖x‖) + χ(‖δ(t)‖) (3.42)
Chapter 3. Application to Axial Flow Compressors 59
∥∥∥∥
∂V
∂x
∥∥∥∥≤ α4(‖x‖) (3.43)
Let X = [x>, z1, z2]> and consider W (X) = V (x) + 1
2z21 + 1
2z22 as a Lyapunov candidate
for system (3.40). Then we have
W =∂V
∂xx+ z1z1 + z2z2
=∂V
∂x[f + g(z1 + u0) + δ] + z1
(
z2 + u1 −A− ∂u0
∂xδ
)
+ z2
[
v −B − ∂u1
∂xδ − ∂u1
∂s1(z2 + u1)
]
Choose u1 and v as follows
u1 = −c1z1 −∂V
∂xg + A+K1
v = −c2z2 − z1 +B +∂u1
∂s1(z2 + u1) +K2
4= π(x, z) (3.44)
where c1, c2 > 0, and K1 and K2 will be defined later. Then we have
W ≤− α3(‖x‖) + χ(‖δ(t)‖)
− c1z21 + z1K1 − z1
∂u0
∂xδ
− c2z22 + z2K2 − z2
∂u1
∂xδ
≤− α3(‖x‖) + χ(‖δ(t)‖)
− c1z21 + z1K1 + z2
1
∥∥∥∥
∂u0
∂x
∥∥∥∥
2
+1
4‖δ‖2
− c2z22 + z2K2 + z2
2
∥∥∥∥
∂u1
∂x
∥∥∥∥
2
+1
4‖δ‖2.
Choosing K1 = −z1∥∥∥∂u0∂x
∥∥∥
2
and K2 = −z2∥∥∥∂u1∂x
∥∥∥
2
, we have
W ≤ −α3(‖x‖) − c1z21 − c2z
22 + χ(‖δ(t)‖) +
1
2‖δ‖2
which implies that W ≤ −π
∥∥∥∥∥∥∥
x
z
∥∥∥∥∥∥∥
+χ1(‖δ‖) for some K∞ function π andK function
χ1.
Chapter 3. Application to Axial Flow Compressors 60
3.2.5 Observability Set
We have that Y = F1(O) = y1e ∈ R
3, z ∈ R2 | y1
e,2 >1β2 (−1 − z1). Define set C as
C =
[y1e>, z>]> ∈ R
5 | y1e,1 ∈ [a1, b1], y
1e,2 ∈
[a2 − z1β2
,b2 − z1β2
]
,
y1e,3 ∈
[1
β2(−z2 + a3),
1
β2(−z2 + b3)
]
, z1 ∈ [a4, b4], z2 ∈ [a5, b5]
,
where ai’s and bi’s are scalars such that ai, bi ∈ R , ai < bi, i = 1, . . . , 5. It can be verified
that the for all a2 > −1, C ⊂ Y . For requirement (ii) in assumption A3 to be satisfied,
each slice C z should be convex. Here the slices are parallelpipeds in R3. The union of all
slices is the set
⋃
z∈R2
C z =
ye ∈ R3 | ye,1 ∈ [a1, b1], ye,2 ∈
[a2 − b4β2
,b2 − a4
β2
]
,
ye,3 ∈[
1
β2(−b5 + a3),
1
β2(−a5 + b3)
]
,
(3.45)
which is compact and satisfying requirement (iv).
The vectors Nyeand Nz are given by
Nye(ye, z) =
[1, 0, 0]> ye,1 = b1
[−1, 0, 0]> ye,1 = a1
[0, 1, 0]> ye,2 = 1β2 (b2 − z1)
[0,−1, 0]> ye,2 = 1β2 (a2 − z1)
[0, 0, 1]> ye,3 = 1β2 (−z2 + b3)
[0, 0,−1]> ye,3 = 1β2 (−z2 + a3),
(3.46)
Chapter 3. Application to Axial Flow Compressors 61
Nz(ye, z) =
[0, 0]> ye,1 = b1 or ye,1 = a1
[1/β2, 0]> ye,2 = 1β2 (b2 − z1)
[−1/β2, 0]> ye,2 = 1β2 (a2 − z1)
[0, 1/β2] > ye,3 = 1β2 (b3 − z2)
[0,−1/β2] > ye,3 = 1β2 (a3 − z2)
[1, 0]> z1 = b4
[−1, 0]> z1 = a4
[0, 1]> z2 = b5
[0,−1]> z2 = a5.
(3.47)
Notice that Nyenever vanishes on any slice C z, which is the condition of the requirement
(iii). Now we have to use the Lyapunov function V to find the largest value of c2 such
that Ωc2 ⊂ O and then pick some values for ai and bi, i = 1, . . . , 5 such that a2 > −1 and
F1(Ωc2) ⊂ C. A more practical way to design C is running a number of simulations for
closed loop system under state feedback for some initial conditions (x(0), z(0)) and finding
upper and lower bounds of ψ(t), φ(t), −ψ(t) − 3/2φ2(t) − 1/2φ3(t) − 3R(t)φ(t) − 3R(t),
z1(t), z2(t), which in turn will give us the values of ai, bi, i = 1, . . . , 5, respectively. By
doing that, we found that whenever
[x(0)>, z(0)>]> ∈ Ω04=[x(0)>, z(0)>]> ∈ R
5 |R ∈ [0, 0.1], φ ∈ [−0.1, 0.1],
ψ ∈ [−0.5, 0.5], z1 ∈ [−0.1, 0.1], z2 ∈ [−0.1, 0.1] ,(3.48)
we have that a1 = −1.15, b1 = 0.5, a2 = −0.3, b2 = −0.1, a3 = −0.75, b3 = 0.4, a4 = −2,
b4 = 7, a5 = −70, b5 = 250.
Chapter 3. Application to Axial Flow Compressors 62
3.2.6 Observer Design
Now we are ready to design observer (2.49) for MG3. Letting xP = [RP , φP , ψP ]>,
f = f(xP , z, y) = [f1, f2, f3]> is given by
f1 = − σ(RP )2 − σRP (2φP + (φP )2)
− l1/ρ+ β2l2/ρ2(3φP + 3RP + 3/2(φP )2) + β2l3/ρ
3
3(1 + φP )(ψ − ψP )
f2 = − ψP − 3/2 (φP )2 − 1/2 (φP )3 − 3RP φP − 3RP + β2l2/ρ2(ψ − ψP )
f3 = − z1 − φP
β2+ l1/ρ(ψ − ψP )
f together with Nyeand Nz in (3.46) and (3.47) respectively, and H1 in (3.11), concludes
the observer and the dynamic projection. The output feedback controller is given by
v = π(xP , v) in (3.44).
3.2.7 Simulation Results
Here we represent the simulation results for the designed projected observer in the previ-
ous sections when applied to MG3 in the presence of disturbances. We choose k1 = 20.43
and k2 = 4.43 × 104 to fulfill ISS inequalities, and L = [6, 12, 18] so that the associated
polynomial s3 + l1s2 + l2s+ l3 = 0 is Hurwitz. Notice that in all simulations x(0) = 0.
As for the disturbances δ(t) = [δ1(t), δ2(t), δ3(t)]> we choose several cases. Figure 3.8 is
the case when δ(t) = [0.001, 0.01sin(t), 0.002cos(t)]> affects the system. The simulation
shows that the output feedback response follows the state feedback one. In Figure 3.9
the system is subjected to constant disturbances δ(t) = [0.003, 0.001, 0.002]> and the
results are shown for two different values of ρ. It can be seen that by decreasing ρ,
the convergence rate of the output feedback increases and the error between the output
feedback and state feedback responses decreases. Figure 3.10 shows the system responses
when δ(t) = [0.03e−t, 0.05e−0.5t, 0.02e−t]>, which is the case where ‖δ(t)‖ → 0 as t→ ∞.
Chapter 3. Application to Axial Flow Compressors 63
It shows that both output feedback and state feedback responses approach to the origin
as t→ ∞. In all three cases output feedback adds mild degradation of performance.
In Figure 3.11 phase curves of the system states and their estimates are represented
when δ(t) = [0.001, 0.01, 0.002]> and ρ = 0.1. In this figure one can observe the effect
of projection on the peaking phenomenon in the observer states, and it can be seen that
the trajectories approach a neighborhood of the origin.
In the simulation for this example we encountered large peaks in z2 = u. This problem
along with the difficulty of finding an appropriate projection set are the reasons that our
initial condition set (3.48) is very conservative.
3.3 Concluding Remarks
In this chapter we applied the techniques we developed in Chapter 2 to the MG3 model
and our simulation results confirmed our theoretical predictions in the previous chapter.
The major difficulty of this technique is finding an appropriate projection set C in (ye, z)
coordinates when its dimension n is greater than 3, as we can see in the MG3 example.
It appears that this particular problem needs further investigation which may lead to a
systematic numerical technique in finding projection sets.
Chapter 3. Application to Axial Flow Compressors 64
0 10 20 300
0.02
0.04
0.06
0.08
0.1
t
R
0 10 20 30−0.1
−0.08
−0.06
−0.04
−0.02
0
0.02
t
φ
0 5 10 15 20 25 30−0.4
−0.3
−0.2
−0.1
0
0.1
t
ψ
Output feedbackState feedback
Figure 3.8: Sinusoidal disturbances (ρ = 1/50 in output feedback).
Chapter 3. Application to Axial Flow Compressors 65
0 10 20 300.02
0.04
0.06
0.08
0.1
0.12
0.14
0.16
t
R
0 10 20 30−0.2
−0.15
−0.1
−0.05
0
t
φ
0 5 10 15 20 25 30−0.6
−0.4
−0.2
0
0.2
t
ψ
state feedbackρ=.2ρ=.01
Figure 3.9: Constant disturbances with different values of ρ.
Chapter 3. Application to Axial Flow Compressors 66
0 10 20 300
0.02
0.04
0.06
0.08
0.1
t
R
0 10 20 30−0.05
−0.04
−0.03
−0.02
−0.01
0
0.01
t
φ
0 5 10 15 20 25 30−0.25
−0.2
−0.15
−0.1
−0.05
0
t
ψ
Output feedbackState feedback
Figure 3.10: Decaying exponential disturbances (ρ = 180
in output feedback).
Chapter 3. Application to Axial Flow Compressors 67
−0.15−0.1
−0.050
0.050.1
0.150.2
0.25
−0.2
−0.1
0
0.1
0.2
−0.4
−0.2
0
0.2
0.4
Rφ
ψ
output feedbackestimatesinitial conditionorigin
Figure 3.11: Peaking phenomenon.
Chapter 4
Observability of Systems with
Uncertainty
In this chapter we apply the observer we developed in Chapter 2 to the MG3 model with
one uncertain parameter and explain the theoretical problem that is encountered when
uncertainties appear in systems in such a way that the affect the observability map. Then
we produce a simple example to show that our observer generates desirable results when
uncertainties do not affect the observability map.
4.1 MG3 with Uncertainty
In Chapter 3 we used an MG3 model without any uncertain parameters. Here we use
the following model,
Φ = −Ψ + ΨC(Φ, θ) − 3ΦR
Ψ =1
β2(Φ − ΦT )
R = σR(1 − Φ2 − R), R(0) ≥ 0
(4.1)
where the difference from the model (3.7) is the presence of the uncertain parameter θ,
which can be considered as an uncertainty in the compressor model ΨC(Φ) = ΦC0 + 1 +
68
Chapter 4. Observability of Systems with Uncertainty 69
θΦ2 − 12Φ3. Notice that the technique used in Chapter 3 copes with small uncertainties
in the constant term ΦC0 + 1. A more general case of uncertainties will be discussed in
Remark 7. Note that here we do not consider time-varying disturbances.
4.2 Controller Design
Following the procedure we outlined in Chapter 3 we first find a controller for the extended
system (system with two integrators at the input side). Shifting the origin to the desired
equilibrium Re = 0,Φe = 1,Ψe = ΨC(Φe) = ΨC0+ 2 with the change of variables
φ = Φ − 1, ψ = Ψ − ΨC0− 2, we have
R = −σR2 − σR(2φ+ φ2)4= f1(x)
φ = −ψ − θφ2 − 1
2φ3 − 3Rφ− 3R = f2(x, θ)
ψ = − 1
β2(u− 1 − φ).
(4.2)
where x = (R, φ, ψ), and ΦT is considered as the control input u. Augmenting the system
with two integrators we have
R = f1 (4.3)
φ = −ψ + w1θ +B (4.4)
ψ =1
β2(1 + φ− ξ1) (4.5)
ξ1 = ξ2 (4.6)
ξ2 = v (4.7)
where B = 12φ3 − 3Rφ − 3R, w1 = φ2. We design the controller in the following steps
using the tuning function approach outlined in [12].
Chapter 4. Observability of Systems with Uncertainty 70
Step 1. First we consider 4.3-4.4. Using the change of variables
z1 = φ (4.8)
z2 = ψ − α1 (4.9)
where α1 = c1φ+ w1θ, c1 > 0, and θ is an estimate of θ, we have
z1 = −z2 − c1z1 + w1(θ) +B
where θ = θ − θ. Using
V1 = CR+1
2z21 +
1
2z22 +
1
2γθ2 C > 0, γ > 0
as a Lyapunov candidate, letting τ1 = z1w1, and noticing that ˙θ = − ˙θ, we have that
V1 = − C(σR2 − σR(2φ+ φ2)) − c1φ2 − φ4/4 − 3Rφ2 − 3Rφ
− z1z2 + θ(τ1 − ˙θ/γ).
Using Young’s inequality we have
V1 ≤− CσR2 − c1φ2 + 2Cσ(R2/4 + φ2) + 3R2/4 + 3φ2
− z1z2 + θ(τ1 −1
γ˙θ/γ)
≤−(Cσ
2− 3
4
)
R2 − (c1 − 2Cσ − 3)φ2
− z1z2 + θ(τ1 − ˙θ/γ)
≤− χ(R, φ) − z1z2 + θ(τ1 − ˙θ/γ)
where χ(R, φ) =(Cσ2− 3
4
)R2 + (c1 − 2Cσ − 3)φ2. Choosing c1 > 2Cσ + 3 and Cσ > 3
2
we have that χ > 0. We eliminate z1z2 in the next step. Since ψ is not the control input,
we postpone the choosing of the update law for θ to the last step.
Step 2. Next we consider 4.3-4.5 and the change of variable
z3 = ξ1 − α2. (4.10)
Chapter 4. Observability of Systems with Uncertainty 71
Then we have that
z2 =1
β2(1 + φ− z3 − α2) − α1.
Choosing α2 = 1 + φ− α′2/β
2 we have
z2 = −z3/β2 + α′2 − α1.
After calculating α1 we have
z2 = − z3/β2 + α′
2 −∂α1
∂φ(−ψ +B) + w2θ − w1
˙θ
= − z3/β2 + α′
2 −∂α1
∂φ(−ψ +B) + w2θ + w2θ − w1
˙θ.
where w2 = −∂α1∂φ
w1. Using V2 = V1 + 12z22 as a Lyapunov candidate and letting τ2 =
z2w2 + τ2 we have that
V2 ≤− χ− z1z2 + z2
(
−z3/β2 + α′2 −
∂α1
∂φ(−ψ +B) + w2θ − w1
˙θ
)
+ (τ2 − ˙θ/γ)θ.
Choosing α′2 as follows
α′2 = z1 +
∂α1
∂φ(−ψ +B) − w2θ +
∂α1
∂θγτ2 − c2z2
where c2 > 0, we have that
V2 ≤ −χ− c2z22 −
1
β2z2z3 + z2
∂α1
∂θ(γτ2 − ˙
θ) + θ(τ2 − ˙θ)
and
z2 = −z3/β2 + z1 + w2θ +∂α
∂θ(γτ2 − ˙
θ).
Step 3. In this step we choose
z4 = ξ2 − α2. (4.11)
Chapter 4. Observability of Systems with Uncertainty 72
Therefore we have
z3 =z4 + α3 − α2
=z4 + α3 + F + w3θ + w3θ −∂α2
∂θ˙θ
where F = −∂α2∂R
f1 − ∂α2∂φ
(−ψ+B)− ∂α2∂ψ
(1 + φ− ξ1)/β2 and w3 = −∂α2
∂ψw1. Choosing
V3 = V2 + 12z23 as Lyapunov candidate and α3 = −c3z3 − F −w3θ+ ∂α2
∂θγτ3 + z2/β
2 + ν3
we have
V3 ≤ −χ− c2z22 − c3z
23 + z3z4 + θ(τ3 − ˙
θ)
(
z2∂α1
∂θ+ z3
∂α2
∂θ
)
(γτ3 − ˙θ)
where τ3 = τ2 + z3w3, and we set ν3 = γz2w3∂α1
∂θHere we used that fact that
˙θ − γτ2 =
˙θ − γτ3 + γz3w3. The time derivative of z3 has the following form
z3 = z4 − c3z3 + w3θ +∂α2
∂θ(γτ3 − ˙
θ) + z2/β2 + ν3.
Step 4. In the last step of this recursive procedure we determine a control law for v.
From (4.11) we have that
z4 =v − α3
=v −G+ w4θ + w4θ −∂α3
∂θ
˙θ
where G = ∂α3∂R
f1 − ∂α3∂φ
(−ψ +B) − ∂α3∂ψ
(1 + φ− ξ1)/β2 − ∂α3
∂ξ1ξ2 and w4 = ∂α3
∂φw1.
Using V4 = V3 + 12z24 as the Lyapunov candidate and choosing
v = −G− w4θ +∂α3
∂θγτ4 − z3 − c4z4 + ν4 (4.12)
where τ4 = τ3 + z4w4, we have
V4 ≤ −χ4 − c2z22 − c3z
23 − c4z
24
Chapter 4. Observability of Systems with Uncertainty 73
0 20 40 60 80 1000
0.005
0.01
0.015
0.02
t
R
0 20 40 60 80 100−5
0
5
10
15
20x 10
−3
t
φ
0 10 20 30 40 50 60 70 80 90 100−0.06
−0.04
−0.02
0
0.02
0.04
0.06
t
ψ
Figure 4.1: MG3 with adaptive controller.
where we use the fact that γ(τ3 − τ4) = −γz4w4 and the following
ν4 = Πγw4
Π = z2∂α1
∂θ+ z3
∂α2
∂θ˙θ = γτ4. (4.13)
We use (4.13) as the update law for the uncertain parameter θ.
Figure 4.1 represents (R, φ, ψ) when controller (4.12) is applied which shows that the
system state tend to zero as t→ ∞.
One drawback of tuning function method is that as the number of integrators increases,
the complexity of the controller grows. Figure 4.1 shows R and ψ vary more slowly com-
paring with φ. This is due to the function α1 in the first step of the controller design.
Chapter 4. Observability of Systems with Uncertainty 74
We can improve this problem by choosing a more complicated function α1 at the first
step, but at the expense of a more complicated final controller v. Due to this complexity,
as it can be seen in the simulation results, we were forced to use small initial conditions.
Remark 7: When there is more than one uncertain parameter in equation (4.4), a more
general controller can be designed using the same step-by-step procedure. Suppose (4.4)
is written as
φ = f2 + w1>θ (4.14)
where θ is a vector of unknown parameters. To solve this problem we can start with
V1 = 12z21 + 1
2θ>Γ−1θ, where Γ is the adaptation gain matrix and is positive definite. At
the end the update law would be˙θ = Γτ(x, θ).
As another case which is more general, the authors in [24] introduce structured parametric
uncertainty to the system (3.7) of the form ∆i as follows
R = f1 + ∆1(R, φ)
φ = f2 + ∆2(R, φ)
ψ =1
β2(Φ − ΦT ),
but their controller is full state feedback. In other words they assume that all the states
can be measured.
4.3 Observability Map
Consider system (4.2) as follows
R = f1(x)
φ = f2(x, θ)
ψ = k(u− 1 − φ).
(4.15)
Chapter 4. Observability of Systems with Uncertainty 75
where x = (R, φ, ψ), and k = − 1β2 . Having y = ψ = x3 as the output, the observability
map for this system is
ye =
y
y
y
=
x3
k(1 + x2) − ku
kf2(x, θ) − ku
4= H(x, z, θ). (4.16)
Here we have a major theoretical problem which is the dependence of H on θ, an un-
known parameter. This problem makes the inverse H−1 unknown, which in our observer
(2.12) in Chapter 2 translates to an undefined[∂H∂x
]−1
. Our attempt to apply the pro-
jected observer to this example has failed so far, and it seems that this problem needs a
substantial investigation which is beyond the scope of this thesis.
But the projected observer in Chapter 2 can be applied to a class of systems with uncer-
tainty which will be discussed in the next section.
4.4 Uncertainty and Observability Map
Consider the following system
x = f(x, u, θ)
y = h(x, u)
(4.17)
where x ∈ Rn, θ ∈ R
p is a vector of unknown parameters, u, y ∈ R, f and h are known
smooth functions and f(0, 0) = 0. Applying the map (2.2) and using the notation in
Chapter 2 we have
ye =
y
y
...
y(n−1)
=
h(x, u)
φ1(x, z, θ)
...
φn−1(x, z, θ)
(4.18)
Chapter 4. Observability of Systems with Uncertainty 76
where z = (u, u, u, . . . , u(nu−1)), θ = (θ, θ, . . . , θ(nθ)), 0 ≤ nu ≤ n and 0 ≤ nθ ≤ n − 1
are integers. The appearance of the unknown term θ in H prevents us from applying
the observer that we developed in Chapter 2. Therefore we need some assumption to
overcome this problem.
Assumption A4(Observability and Uncertainty): Assume that the plant has the
following property
∂φi∂θ
= 0, for i = 0, . . . , n− 1. (4.19)
Therefore (4.18) can be written as ye = H(x, z).
One class of systems that has this property is
x1 = f1(x1, x2)
x2 = f2(x1, x2, x3)
...
xi = fi(x1, . . . , xi+1)
...
xn = fn(x, u, θ)
y = h(x1).
(4.20)
For a further discussion on this problem one can refer to [25].
The next two examples are systems that have the structure of the form (4.20). The first
example is a system that is observable everywhere while the second one does not have
this property.
Chapter 4. Observability of Systems with Uncertainty 77
0 0.5 1 1.5 2 2.5 3−0.2
0
0.2
0.4
0.6
0.8
1
1.2
t
x1
0 0.5 1 1.5 2 2.5 3−4
−2
0
2
4
t
x2
system trajectoryestimates
Figure 4.2: x and its estimate, Example 4.4.1.
Example 4.4.1 Consider the system x = f(x, θ) as follows
x1 = 2x1 + x2
x2 = u+ wθ
y = x1
(4.21)
where w = x21 + .5x2 and θ is an unknown parameter. This system has two interesting
features. First, θ dose not appear in the observability map as we will show shortly.
Second, the control input u does not appear in the observability map and we do not need
to use integrators at the input side.
Using adaptive backstepping as in Section 4.2 we find the following state feedback con-
Chapter 4. Observability of Systems with Uncertainty 78
−4 −3 −2 −1 0 1 2 3 4−4
−3
−2
−1
0
1
2
3
4
x1
x2
initial conditionoriginsystem trajectoryestimates
Figure 4.3: Projection effect, Example 4.4.1.
Chapter 4. Observability of Systems with Uncertainty 79
troller
u = −z1 − c2z2 + α1 − wθ (4.22)
z2 = x2 − α1
z1 = x1
α1 = −c1z1
where c1 > 2, c2 > 0 and
˙θ = γz2w (4.23)
where γ > 0 is the update law for the unknown parameter. Following the procedure to
design an observer we have (notice that, due to the absence of disturbance, H = H1)
ye =
y
y
=
x1
x2
= H(x) (4.24)
∂H
∂x=
1 0
0 1
L = [l1 l2]> = [2 1]>
E =
ρ 0
0 ρ2
f(x, u, y) =
x1
u(x) + w(x) + θ0
+ E−1L(y − x1)
where θ0 is the nominal value of θ. As for the projection surface C we use the surface of
a circle of radius√
14 which is x21 + x2
2 − 14 = 0. Next we have to define the projection
elements as follows
Nye= [xP1 xP2 ]> , Nz = 0
∂H
∂z= 0. (4.25)
Chapter 4. Observability of Systems with Uncertainty 80
The simulation results are depicted in Figures 4.2-4.3. In this simulation we have chosen
ρ = 0.02, θ0 = 1 and θ = 3. In Figure 4.3 one can see the effect of projection on circle
surface for x. Notice that the x(0) = 0.
44
Example 4.4.2 Consider the system x = f(x, θ) as follows
x1 = x21 + 2x2
x2 = u+ wθ
y = (x1 − 1)2
(4.26)
where w = x22ex1 and θ is an unknown parameter.
First we design a state feedback controller. Using adaptive backstepping and considering
V = 12z21 + 1
2z22 + 1
2γas a Lyapunov candidate, we find the following state feedback
controller
u = −2z1 − c2z2 + α1 − wθ (4.27)
z2 = x2 − α1
z1 = x1
α1 = −c1z1 − x21
where c1 > 0, c2 > 0 and
˙θ = γz2w (4.28)
with γ > 0 is the update law for the unknown parameter. Using this controller we have
V ≤ −c1z21 − c2z
22 .
Following the procedure to design an observer we have (notice that, due to the absence
Chapter 4. Observability of Systems with Uncertainty 81
of disturbance, H = H1)
ye =
y
y
=
(x1 − 1)2
2(x21 + 2x2)(x1 − 1)
= H(x) (4.29)
∂H
∂x=
2(x1 − 1) 0
4(x21 − x1) + 2(x2
1 + 2x2) 4(x1 − 1)
.
(4.30)
Notice that nu = 0 and H is invertible for x1 < 1. Therefore we have to choose the set C
in ye coordinates such that it contains H(0, 0) and ye,1 > 1. For that purpose we choose
C to be the following set
C =(ye ∈ R
2|(ye,1 − 1)2/.9025 + y2e,2/4 ≤ 1
which is an ellipse. Next we have to define the observer and the projection elements as
follows
L = [l1 l2]> = [2 1]>
E =
ρ 0
0 ρ2
f(x, u, y) =
x1
u(x) + w(x) + θ0
+
[∂H
∂x
]−1
E−1L(y − (x1 − 1)2)
Nye= ye + [−1/.9025 1/4]> , Nz = 0
∂H
∂z= 0.
The simulation results are depicted in Figures 4.4, 4.5 and 4.6. In this simulation we
have chosen ρ = 0.02, θ0 = 1 and θ = 3. Notice that the x(0) = 0. Figure 4.5 shows the
effect of the projection on x in x coordinates, while Figure 4.6 shows the effect of the
projection on the estimates when the hit the ellipse surface (ye,1 − 1)2/.9025+ y2e,2/4 = 1
in ye coordinates.
Chapter 4. Observability of Systems with Uncertainty 82
0 0.5 1 1.5 2 2.5−0.1
−0.05
0
0.05
0.1
0.15
0.2
0.25
t
x1
0 0.5 1 1.5 2 2.5−0.8
−0.6
−0.4
−0.2
0
0.2
0.4
0.6
t
x2
state feedbackoutput feedbackestimats
Figure 4.4: x and its estimate, Example 4.4.2.
Figures 4.7 and 4.8 represent the simulation results when disturbance
δ(t) = [0.4sin(t) 0.3cos(t)]> is added to the system in the form of x = f(x, θ)+δ(t). This
is an example when there are disturbances and uncertainty in system, in other words a
combination of the results of Chapter 2 and this chapter. Figure 4.8 shows that, because
of the existence of disturbance, the trajectories enter a neighborhood of the origin, which
is an example of the ultimate boundedness result in Chapter 2.
44
Chapter 4. Observability of Systems with Uncertainty 83
−0.2 −0.15 −0.1 −0.05 0 0.05 0.1 0.15 0.2
−0.6
−0.4
−0.2
0
0.2
0.4
x1
x2
state feedbackoutput feedbackestimatsinitial conditionorigin
Figure 4.5: Projection effect in x coordinates, Example 4.4.2.
Chapter 4. Observability of Systems with Uncertainty 84
0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2−2.5
−2
−1.5
−1
−0.5
0
0.5
1
1.5
2
2.5
ye1
ye
2
initial conditionoriginoutput feedbackestimats
Figure 4.6: Projection effect in ye coordinates, Example 4.4.2.
Chapter 4. Observability of Systems with Uncertainty 85
0 1 2 3 4 5 6 7 8 9 10−0.1
−0.05
0
0.05
0.1
0.15
0.2
0.25
t
x1
0 1 2 3 4 5 6 7 8 9 10−0.8
−0.6
−0.4
−0.2
0
0.2
0.4
0.6
t
x2
state feedbackoutput feedbackestimats
Figure 4.7: Disturbance and uncertainty, Example 4.4.2.
Chapter 4. Observability of Systems with Uncertainty 86
−0.2 −0.15 −0.1 −0.05 0 0.05 0.1 0.15 0.2
−0.6
−0.4
−0.2
0
0.2
0.4
x1
x2
state feedbackoutput feedbackestimatsinitial conditionorigin
Figure 4.8: Disturbance and uncertainty, projection effect, Example 4.4.2.
Chapter 5
Conclusion
Our approach to the output feedback control of systems affected by disturbances in
Chapter 2 allows us to use available state feedback control design techniques. However,
in practice, during the design step one may encounter some limitations and difficulties
as we did in Chapters 3 and 4.
First, our method requires adding integrators at the input side. This complicates the
state feedback design and generally leads to a complex expression for the state feedback
function in the last integrator. Second, it appears that finding an appropriate set C for
the projection may be a difficult task and sometimes it may yields a very conservative
set which limits our control effort and/or our ability to use the current simulation tools
such as MATLAB. For example, due to our choice of set C in Chapter 3 for MG3, our
choice of initial conditions for simulation is limited to small values. Third, in Chapter 3,
our technique relies on the perfect knowledge of the compressor characteristic which is
not a realistic assumption. Our attempts to add some limited kind of uncertainties in the
compressor characteristic hasve failed as we discussed in Chapter 4. Additionally, we just
solved the problem for a limited kind of disturbances with a conservative assumption.
Some points that we mentioned throughout this thesis and the limitations that we just
mentioned set some possible directions for future research such as, finding a Lyapunov
87
Chapter 5. Conclusion 88
or non-Lyapunov based proof for the Conjecture in Chapter 2, a possible proof for the
combination of the results in Chapters 2 and 4, modifying the observer to work with
systems with a broader range of uncertainties and disturbances, developing numerical
methods to automate the design of C, and developing some estimation methods other
than backstepping for the control input. Additionally, it would be interesting to find
a possibly modified version of this method for systems that are unobservable at some
individual points or even at the origin.
Bibliography
[1] F. Esfandiari, H. Khalil, Output feedback stabilization of fully linearizable systems,
International Journal of Control, 56(5):1007-1037, 1992. 1.1.1
[2] F. Esfandiari, H. Khalil, Semiglobal stabilization of a class of nonlinear systems
using output feedback, IEEE Transactions on Automatic Control, 38(9):1412-1415,
1993 1.1.1
[3] Z. Lin, A. Saberi, Robust semiglobal stabilization of minimum-phase input-output
linearizable systems via partial state and out put feedback, IEEE Transactions on
Automatic Control, 40(6):1029-1041, 1995. 1.1.1
[4] M. Jankovic, Adaptive output feedback control of nonlinear feedback linearizable sys-
tems, International Journal of Adaptive Control and Signal Processing, 10:1-18,
1996. 1.1.1
[5] M. Mahmoud, H. Khalil, Asymptotic regulation of minimum phase nonlinear systems
using output feedback, IEEE Transactions on Automatic Control, 41(10):1402-1412,
1996. 1.1.1
[6] M. Mahmoud, H. Khalil, Robust control for a nonlinear servomechanism problem,
International Journal of Control, 66(6):779-802, 1997. 1.1.1
89
Bibliography 90
[7] A. Attasi, H. Khalil, A seperation principle for the stabilization of a class of nonlinear
systems, IEEE Transactions on Automatic Control, 44:1672-1678, September 1999.
1.1.1, 2.2
[8] A. Tornambe, Output feedback stabilization of a class of non-minimum phase non-
linear systems, Systems and Control letters, 19:193-204, 1992. 2.2
[9] A. Teel, L. Praly, Global stabilizability and observability imply semi-global stabiliz-
ability by output feedback, Systems and Control letters, 22:313-325, 1994. 1.1.1,
2.2
[10] E.D. Sontag, Remarks on Stabilization and input to state stability, Proceedinds of the
IEEE Conference on Decision and Control, (Tampa, FL), pp. 1376-1378, December
1989. 3
[11] A. Isidori, Nonlinear Control Systems, London: Springer-Verlag, Third ed., 1995. 3
[12] M. Krstic, I. Kanellakopoulos, and P. Kokotovic, Nonlinear and Adaptive Control
Design, NY: John Wiley & Sons, Inc., 1995. 2, 3, 3.2.2, 4.2
[13] J. Kurzweil, On the inverse of Lyapunov’s second theorem on stability of motion,
Americam Mathematical Society Translation, series 2, 24:19-77, 1956. 2.3.1
[14] H.K. Khalil , Nonlinear Systems, NJ: Printice Hall, Inc., Second ed., 1996. 3, 2.2,
2.3.1, 2.4
[15] P. Ioannou and J. Sun , Stable and Robust Adaptive Control, Englewood Cliffs, NJ:
Printice Hall, Second ed., 1995. 2.3
[16] M. Maggiore, K. Passino, A Separation Principle for a Class of Non-Uniformly
Completely Observable Systems, IEEE Transactions on Automatic Control, 48(7),
2003. 1, 1.1.1, 1.2, 2, 2.2, 6, 2.4
Bibliography 91
[17] T.P. Hynes, E.M. Greitzer, A method for assessing effects of circunferential flow
distortion on compressor stability, Journal of Turbomachinery, 109:371-379, 1987.
1.1.2
[18] J.C. DeLaat, R.D. Southwick, G.W. Gallops, High stability engine control (HIS-
TEC), In: Proc. AIAA/SAE/ASME/ASEE 32nd Joint Propulsion Conf., Lake
Buena Vista, FL. NASA TM 107272, AIAA-96-2586. 1.1.2
[19] J.S. Simon, L. Valavani, A Lyapunov based nonlinear control scheme for stabilizing
a basic compression system using a closed couple control valve, In: Proc. American
Control Conference, pp. 2398-2406, 1991. 1.1.2, 3.1.6
[20] W.M. Haddad, J.L. Fausz, V.-S. Chellaboina, A. Leonessa, Nonlinear robust distur-
bance rejection controllers for rotating stall and surge in axial flow compressors, In:
Proc. International Conf. on Control Applications, Hartford, CT,
pp. 767-772, 1997. 1.1.2
[21] S. Baghdadi, J.E. Lueke, Compressor stability analysis, Journal of Fluids Engineer-
ing, 104:242-249, 1982. 1.1.2
[22] C.F. Lorenzo, F.P. Chiaramante and C.M. Mehalic , Determination of compressor
in-stall characteristics from engin surge transients, Journal of Propulsion and Power
4(2): 133-142, 1986. 1
[23] M. Maggiore, K. Passino, A separation principle for Non-UCO systems: the jet
engine surge and stall example, IEEE Transactions on Automatic Control, 48(7),
2003. 1, 1.1.2, 3, 3.2.2
[24] W.M. Haddad, A. Leonessa, V.-S. Chellaboina, J.L. Fausz, Nonlinear robust dis-
turbance rejection controllers for rotating stall and surge in axial flow compressors,
IEEE Transactions on Control Systems Technology, 7(3):391-398, 1999. 1.1.2, 7
Bibliography 92
[25] J.T. Spooner, M. Maggiore, R. Qrdonez, K.M. Passino, Stable Adaptive Control and
Estimation for Nonlinear Systems, Neural and Fuzzy Approximation Techniques,
John Wiley and Sons Innc., New York, 2002. 4.4
[26] F.K. Moore, E.M. Greitzer,A theory of post-stall transients in axial compression
systems, part I, development of equations, Journal of Turbomachinery, 108:68-76,
1986. 1
[27] E.H. Abed, P.K. Houpt, W.M. Hosny, Bifurcation analysis of surge and rotating stall
in axial flow compressors, Journal of Turbomachinery, 115:817-824, 1993. 1.1.2
[28] O.O. Badmus, S. Chowdhury, C.N. Nett, Nonlinear control of surge in axial com-
pression systems, Automatica, 32(1):59-70, 1996. 1.1.2
[29] A.H. Epstein, J.F. Williams, E.M. Greitzer, Active suppression of aerodynamic in-
stabilities in turbomachinery, J. Propulsion, 5:204-211, 1989. 1.1.2
[30] K. Eveker, D. Gysling, C. Nett, O. Sharma, Integrated control of rotating stall and
surge in high speed multistage compression systems, Journal of Turbomachinery,
120(3):440-445, 1998. 1.1.2
[31] M. Krstic, D. Fontaine, P.V. Kokotovic, J.D. Paduano, Useful nonlinearities and
global stabilization of bifurcations in a model of jet engine surge and stall, IEEE
Transactions on Automatic Control, 43(12):1739-1745, 1998. 1.1.2, 3
[32] D. Liaw, E. Abed, Stability analysis and control of rotating stall, In Proceedings of
the IFAC Nonlinear Control Systems Design Symposium, Bordeaux, France, June
1992. 1.1.2
[33] D. Liaw, E. Abed, Active control of compressor stall inception: a bifurcation-
theoretic approach, Automatica, 32(1):109-115, 1996. 1.1.2
Bibliography 93
[34] F.E. McCaughaun, Bifurcation analysis of axial flow compressor stability, SIAM
Journal of Applied Mathematics, 50(5):1232-1253, 1990. 1.1.2
[35] J.D. Paduano, L. Valavani, A. Epstein, E. Greitzer, G.R. Guenette, Modeling for
control of rotating stall, Automatica, 30(9):1357-1373, 1994. 1.1.2
[36] J.J. Wang, M. Krstic, M. Larsen, Control of deep hysteresis aero-engine compressor,
In Proc. American Control Conf., pp. 998-1007, Alburquerque, NM, 1997. 1.1.2
[37] F. Willems, M. Heemels, B. de Jager, A. Stoorvogel, Positive feedback stabilization
of compressor surge, In Proc. 38th Conf. Decision Control, pp. 3259-3264, Phoenix,
AZ, December 1999. 1.1.2
[38] B. de Jager, Rotating stall and surge control: A survey, Proc. of the 34th Conf. on
Decision and Control, pp. 1857-1862, New Orleans, LA, December 1995. 3
[39] T. Gravdhal, O. Egeland, Compressor Surge and Rotating Stall, Modeling and Con-
trol, Advanced Industrial Control, Springer-Verlag London Limited, 1999. 1.1.2,
3
[40] A. Isidori, Nonlinear Control Systems II, Springer-Verlag, London, 1999. 1.1.2